LEDITS++

LEDITS++ was proposed in LEDITS++: Limitless Image Editing using Text-to-Image Models by Manuel Brack, Felix Friedrich, Katharina Kornmeier, Linoy Tsaban, Patrick Schramowski, Kristian Kersting, Apolinário Passos.

The abstract from the paper is:

Text-to-image diffusion models have recently received increasing interest for their astonishing ability to produce high-fidelity images from solely text inputs. Subsequent research efforts aim to exploit and apply their capabilities to real image editing. However, existing image-to-image methods are often inefficient, imprecise, and of limited versatility. They either require time-consuming fine-tuning, deviate unnecessarily strongly from the input image, and/or lack support for multiple, simultaneous edits. To address these issues, we introduce LEDITS++, an efficient yet versatile and precise textual image manipulation technique. LEDITS++'s novel inversion approach requires no tuning nor optimization and produces high-fidelity results with a few diffusion steps. Second, our methodology supports multiple simultaneous edits and is architecture-agnostic. Third, we use a novel implicit masking technique that limits changes to relevant image regions. We propose the novel TEdBench++ benchmark as part of our exhaustive evaluation. Our results demonstrate the capabilities of LEDITS++ and its improvements over previous methods. The project page is available at https://leditsplusplus-project.static.hf.space .

Tip

You can find additional information about LEDITS++ on the project page and try it out in a demo.

Tip

Due to some backward compatability issues with the current diffusers implementation of [~schedulers.DPMSolverMultistepScheduler] this implementation of LEdits++ can no longer guarantee perfect inversion. This issue is unlikely to have any noticeable effects on applied use-cases. However, we provide an alternative implementation that guarantees perfect inversion in a dedicated GitHub repo.

We provide two distinct pipelines based on different pre-trained models.

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion

Bases: DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin

Pipeline for textual image editing using LEDits++ with Stable Diffusion.

This model inherits from [DiffusionPipeline] and builds on the [StableDiffusionPipeline]. Check the superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular device, etc.).

PARAMETER	DESCRIPTION
vae	Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. TYPE: [`AutoencoderKL`]
text_encoder	Frozen text-encoder. Stable Diffusion uses the text portion of CLIP, specifically the clip-vit-large-patch14 variant. TYPE: [`~transformers.CLIPTextModel`]
tokenizer	Tokenizer of class CLIPTokenizer. TYPE: [`~transformers.CLIPTokenizer`]
unet	Conditional U-Net architecture to denoise the encoded image latents. TYPE: [`UNet2DConditionModel`]
scheduler	A scheduler to be used in combination with unet to denoise the encoded image latens. Can be one of [DPMSolverMultistepScheduler] or [DDIMScheduler]. If any other scheduler is passed it will automatically be set to [DPMSolverMultistepScheduler]. TYPE: [`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]
safety_checker	Classification module that estimates whether generated images could be considered offensive or harmful. Please, refer to the model card for details. TYPE: [`StableDiffusionSafetyChecker`]
feature_extractor	Model that extracts features from generated images to be used as inputs for the safety_checker. TYPE: [`~transformers.CLIPImageProcessor`]

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

class LEditsPPPipelineStableDiffusion(
    DiffusionPipeline, TextualInversionLoaderMixin, StableDiffusionLoraLoaderMixin, IPAdapterMixin, FromSingleFileMixin
):
    """
    Pipeline for textual image editing using LEDits++ with Stable Diffusion.

    This model inherits from [`DiffusionPipeline`] and builds on the [`StableDiffusionPipeline`]. Check the superclass
    documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular
    device, etc.).

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`~transformers.CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        tokenizer ([`~transformers.CLIPTokenizer`]):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
            [`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]. If any other scheduler is passed it will
            automatically be set to [`DPMSolverMultistepScheduler`].
        safety_checker ([`StableDiffusionSafetyChecker`]):
            Classification module that estimates whether generated images could be considered offensive or harmful.
            Please, refer to the [model card](https://huggingface.co/CompVis/stable-diffusion-v1-4) for details.
        feature_extractor ([`~transformers.CLIPImageProcessor`]):
            Model that extracts features from generated images to be used as inputs for the `safety_checker`.
    """

    model_cpu_offload_seq = "text_encoder->unet->vae"
    _exclude_from_cpu_offload = ["safety_checker"]
    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: Union[DDIMScheduler, DPMSolverMultistepScheduler],
        safety_checker: StableDiffusionSafetyChecker,
        feature_extractor: CLIPImageProcessor,
        requires_safety_checker: bool = True,
    ):
        super().__init__()

        if not isinstance(scheduler, DDIMScheduler) and not isinstance(scheduler, DPMSolverMultistepScheduler):
            scheduler = DPMSolverMultistepScheduler.from_config(
                scheduler.config, algorithm_type="sde-dpmsolver++", solver_order=2
            )
            logger.warning(
                "This pipeline only supports DDIMScheduler and DPMSolverMultistepScheduler. "
                "The scheduler has been changed to DPMSolverMultistepScheduler."
            )

        if scheduler is not None and getattr(scheduler.config, "steps_offset", 1) != 1:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
                "to update the config accordingly as leaving `steps_offset` might led to incorrect results"
                " in future versions. If you have downloaded this checkpoint from the Hugging Face Hub,"
                " it would be very nice if you could open a Pull request for the `scheduler/scheduler_config.json`"
                " file"
            )
            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["steps_offset"] = 1
            scheduler._internal_dict = FrozenDict(new_config)

        if scheduler is not None and getattr(scheduler.config, "clip_sample", False) is True:
            deprecation_message = (
                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
                " config accordingly as not setting `clip_sample` in the config might lead to incorrect results in"
                " future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it would be very"
                " nice if you could open a Pull request for the `scheduler/scheduler_config.json` file"
            )
            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(scheduler.config)
            new_config["clip_sample"] = False
            scheduler._internal_dict = FrozenDict(new_config)

        if safety_checker is None and requires_safety_checker:
            logger.warning(
                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
            )

        if safety_checker is not None and feature_extractor is None:
            raise ValueError(
                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
            )

        is_unet_version_less_0_9_0 = (
            unet is not None
            and hasattr(unet.config, "_diffusers_version")
            and version.parse(version.parse(unet.config._diffusers_version).base_version) < version.parse("0.9.0.dev0")
        )
        is_unet_sample_size_less_64 = (
            unet is not None and hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
        )
        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
            deprecation_message = (
                "The configuration file of the unet has set the default `sample_size` to smaller than"
                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
                " in the config might lead to incorrect results in future versions. If you have downloaded this"
                " checkpoint from the Hugging Face Hub, it would be very nice if you could open a Pull request for"
                " the `unet/config.json` file"
            )
            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
            new_config = dict(unet.config)
            new_config["sample_size"] = 64
            unet._internal_dict = FrozenDict(new_config)

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            unet=unet,
            scheduler=scheduler,
            safety_checker=safety_checker,
            feature_extractor=feature_extractor,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
        self.register_to_config(requires_safety_checker=requires_safety_checker)

        self.inversion_steps = None

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.run_safety_checker
    def run_safety_checker(self, image, dtype):
        if self.safety_checker is None:
            has_nsfw_concept = None
        else:
            if ops.is_tensor(image):
                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
            else:
                feature_extractor_input = self.image_processor.numpy_to_pil(image)
            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="np")
            image, has_nsfw_concept = self.safety_checker(
                images=image, clip_input=ms.tensor(safety_checker_input.pixel_values).to(dtype)
            )
        return image, has_nsfw_concept

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
    def decode_latents(self, latents):
        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)

        latents = 1 / self.vae.config.scaling_factor * latents
        image = self.vae.decode(latents, return_dict=False)[0]
        image = (image / 2 + 0.5).clamp(0, 1)
        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
        image = image.permute(0, 2, 3, 1).float().numpy()
        return image

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, eta, generator=None):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.check_inputs
    def check_inputs(
        self,
        negative_prompt=None,
        editing_prompt_embeddings=None,
        negative_prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
    ):
        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"  # noqa: E501
            )
        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if editing_prompt_embeddings is not None and negative_prompt_embeds is not None:
            if editing_prompt_embeddings.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`editing_prompt_embeddings` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `editing_prompt_embeddings` {editing_prompt_embeddings.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

    # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, latents):
        # shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)

        # if latents.shape != shape:
        #    raise ValueError(f"Unexpected latents shape, got {latents.shape}, expected {shape}")

        latents = latents

        # scale the initial noise by the standard deviation required by the scheduler
        latents = latents * self.scheduler.init_noise_sigma
        return latents

    def prepare_unet(self, attention_store, PnP: bool = False):
        attn_procs = {}
        for name in self.unet.attn_processors.keys():
            if name.startswith("mid_block"):
                place_in_unet = "mid"
            elif name.startswith("up_blocks"):
                place_in_unet = "up"
            elif name.startswith("down_blocks"):
                place_in_unet = "down"
            else:
                continue

            if "attn2" in name and place_in_unet != "mid":
                attn_procs[name] = LEDITSCrossAttnProcessor(
                    attention_store=attention_store,
                    place_in_unet=place_in_unet,
                    pnp=PnP,
                    editing_prompts=self.enabled_editing_prompts,
                )
            else:
                attn_procs[name] = AttnProcessor()

        self.unet.set_attn_processor(attn_procs)

    def encode_prompt(
        self,
        num_images_per_prompt,
        enable_edit_guidance,
        negative_prompt=None,
        editing_prompt=None,
        negative_prompt_embeds: Optional[ms.Tensor] = None,
        editing_prompt_embeds: Optional[ms.Tensor] = None,
        lora_scale: Optional[float] = None,
        clip_skip: Optional[int] = None,
    ):
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            enable_edit_guidance (`bool`):
                whether to perform any editing or reconstruct the input image instead
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            editing_prompt (`str` or `List[str]`, *optional*):
                Editing prompt(s) to be encoded. If not defined, one has to pass `editing_prompt_embeds` instead.
            editing_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
                provided, text embeddings will be generated from `prompt` input argument.
            negative_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            lora_scale (`float`, *optional*):
                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
        """
        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            scale_lora_layers(self.text_encoder, lora_scale)

        batch_size = self.batch_size
        num_edit_tokens = None

        if negative_prompt_embeds is None:
            uncond_tokens: List[str]
            if negative_prompt is None:
                uncond_tokens = [""] * batch_size
            elif isinstance(negative_prompt, str):
                uncond_tokens = [negative_prompt]
            elif batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but exoected"
                    f"{batch_size} based on the input images. Please make sure that passed `negative_prompt` matches"
                    " the batch size of `prompt`."
                )
            else:
                uncond_tokens = negative_prompt

            # textual inversion: procecss multi-vector tokens if necessary
            if isinstance(self, TextualInversionLoaderMixin):
                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

            uncond_input = self.tokenizer(
                uncond_tokens,
                padding="max_length",
                max_length=self.tokenizer.model_max_length,
                truncation=True,
                return_tensors="np",
            )

            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
                attention_mask = ms.tensor(uncond_input.attention_mask)
            else:
                attention_mask = None

            negative_prompt_embeds = self.text_encoder(
                ms.tensor(uncond_input.input_ids),
                attention_mask=attention_mask,
            )
            negative_prompt_embeds = negative_prompt_embeds[0]

        if self.text_encoder is not None:
            prompt_embeds_dtype = self.text_encoder.dtype
        elif self.unet is not None:
            prompt_embeds_dtype = self.unet.dtype
        else:
            prompt_embeds_dtype = negative_prompt_embeds.dtype

        negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype)

        if enable_edit_guidance:
            if editing_prompt_embeds is None:
                # textual inversion: procecss multi-vector tokens if necessary
                # if isinstance(self, TextualInversionLoaderMixin):
                #    prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
                if isinstance(editing_prompt, str):
                    editing_prompt = [editing_prompt]

                max_length = negative_prompt_embeds.shape[1]
                text_inputs = self.tokenizer(
                    [x for item in editing_prompt for x in repeat(item, batch_size)],
                    padding="max_length",
                    max_length=max_length,
                    truncation=True,
                    return_tensors="np",
                    return_length=True,
                )

                num_edit_tokens = text_inputs.length - 2  # not counting startoftext and endoftext
                text_input_ids = text_inputs.input_ids
                untruncated_ids = self.tokenizer(
                    [x for item in editing_prompt for x in repeat(item, batch_size)],
                    padding="longest",
                    return_tensors="np",
                ).input_ids

                if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not np.array_equal(
                    text_input_ids, untruncated_ids
                ):
                    removed_text = self.tokenizer.batch_decode(
                        untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
                    )
                    logger.warning(
                        "The following part of your input was truncated because CLIP can only handle sequences up to"
                        f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                    )

                if (
                    hasattr(self.text_encoder.config, "use_attention_mask")
                    and self.text_encoder.config.use_attention_mask
                ):
                    attention_mask = ms.tensor(text_inputs.attention_mask)
                else:
                    attention_mask = None

                text_input_ids = ms.tensor(text_input_ids)
                if clip_skip is None:
                    editing_prompt_embeds = self.text_encoder(text_input_ids, attention_mask=attention_mask)
                    editing_prompt_embeds = editing_prompt_embeds[0]
                else:
                    editing_prompt_embeds = self.text_encoder(
                        text_input_ids, attention_mask=attention_mask, output_hidden_states=True
                    )
                    # Access the `hidden_states` first, that contains a tuple of
                    # all the hidden states from the encoder layers. Then index into
                    # the tuple to access the hidden states from the desired layer.
                    editing_prompt_embeds = editing_prompt_embeds[-1][-(clip_skip + 1)]
                    # We also need to apply the final LayerNorm here to not mess with the
                    # representations. The `last_hidden_states` that we typically use for
                    # obtaining the final prompt representations passes through the LayerNorm
                    # layer.
                    editing_prompt_embeds = self.text_encoder.text_model.final_layer_norm(editing_prompt_embeds)

            editing_prompt_embeds = editing_prompt_embeds.to(dtype=negative_prompt_embeds.dtype)

            bs_embed_edit, seq_len, _ = editing_prompt_embeds.shape
            editing_prompt_embeds = editing_prompt_embeds.to(dtype=negative_prompt_embeds.dtype)
            editing_prompt_embeds = editing_prompt_embeds.tile((1, num_images_per_prompt, 1))
            editing_prompt_embeds = editing_prompt_embeds.view(bs_embed_edit * num_images_per_prompt, seq_len, -1)

        # get unconditional embeddings for classifier free guidance

        # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
        seq_len = negative_prompt_embeds.shape[1]

        negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype)

        negative_prompt_embeds = negative_prompt_embeds.tile((1, num_images_per_prompt, 1))
        negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        if isinstance(self, StableDiffusionLoraLoaderMixin):
            # Retrieve the original scale by scaling back the LoRA layers
            unscale_lora_layers(self.text_encoder, lora_scale)

        return editing_prompt_embeds, negative_prompt_embeds, num_edit_tokens

    @property
    def guidance_rescale(self):
        return self._guidance_rescale

    @property
    def clip_skip(self):
        return self._clip_skip

    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    def __call__(
        self,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        generator: Optional[Union[np.random.Generator, List[np.random.Generator]]] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = False,
        editing_prompt: Optional[Union[str, List[str]]] = None,
        editing_prompt_embeds: Optional[ms.Tensor] = None,
        negative_prompt_embeds: Optional[ms.Tensor] = None,
        reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
        edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
        edit_warmup_steps: Optional[Union[int, List[int]]] = 0,
        edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
        edit_threshold: Optional[Union[float, List[float]]] = 0.9,
        user_mask: Optional[ms.Tensor] = None,
        sem_guidance: Optional[List[ms.Tensor]] = None,
        use_cross_attn_mask: bool = False,
        use_intersect_mask: bool = True,
        attn_store_steps: Optional[List[int]] = [],
        store_averaged_over_steps: bool = True,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
    ):
        r"""
        The call function to the pipeline for editing. The
        [`~pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.invert`] method has to be called beforehand. Edits will
        always be performed for the last inverted image(s).

        Args:
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
                if `guidance_scale` is less than `1`).
            generator (`np.random.Generator`, *optional*):
                One or a list of [np.random.Generator(s)](https://numpy.org/doc/stable/reference/random/generator.html)
                to make generation deterministic.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] instead of a plain
                tuple.
            editing_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. The image is reconstructed by setting
                `editing_prompt = None`. Guidance direction of prompt should be specified via
                `reverse_editing_direction`.
            editing_prompt_embeds (`ms.Tensor>`, *optional*):
                Pre-computed embeddings to use for guiding the image generation. Guidance direction of embedding should
                be specified via `reverse_editing_direction`.
            negative_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
            reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
                Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
            edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
                Guidance scale for guiding the image generation. If provided as list values should correspond to
                `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 12 of [LEDITS++
                Paper](https://arxiv.org/abs/2301.12247).
            edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
                Number of diffusion steps (for each prompt) for which guidance will not be applied.
            edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
                Number of diffusion steps (for each prompt) after which guidance will no longer be applied.
            edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
                Masking threshold of guidance. Threshold should be proportional to the image region that is modified.
                'edit_threshold' is defined as 'λ' of equation 12 of [LEDITS++
                Paper](https://arxiv.org/abs/2301.12247).
            user_mask (`ms.Tensor`, *optional*):
                User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s
                implicit masks do not meet user preferences.
            sem_guidance (`List[ms.Tensor]`, *optional*):
                List of pre-generated guidance vectors to be applied at generation. Length of the list has to
                correspond to `num_inference_steps`.
            use_cross_attn_mask (`bool`, defaults to `False`):
                Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask
                is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of [LEDITS++
                paper](https://arxiv.org/pdf/2311.16711.pdf).
            use_intersect_mask (`bool`, defaults to `True`):
                Whether the masking term is calculated as intersection of cross-attention masks and masks derived from
                the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate
                are defined as 'M^2' of equation 12 of [LEDITS++ paper](https://arxiv.org/pdf/2311.16711.pdf).
            attn_store_steps (`List[int]`, *optional*):
                Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes.
            store_averaged_over_steps (`bool`, defaults to `True`):
                Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If
                False, attention maps for each step are stores separately. Just for visualization purposes.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            guidance_rescale (`float`, *optional*, defaults to 0.0):
                Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
                using zero terminal SNR.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.

        Examples:

        Returns:
            [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
            returning a tuple, the first element is a list with the generated images, and the second element is a list
            of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
            content, according to the `safety_checker`.
        """

        if self.inversion_steps is None:
            raise ValueError(
                "You need to invert an input image first before calling the pipeline. The `invert` method has to be called beforehand. Edits will always be performed for the last inverted image(s)."  # noqa: E501
            )

        eta = self.eta
        num_images_per_prompt = 1
        latents = self.init_latents

        zs = self.zs
        self.scheduler.set_timesteps(len(self.scheduler.timesteps))

        if use_intersect_mask:
            use_cross_attn_mask = True

        if use_cross_attn_mask:
            self.smoothing = LeditsGaussianSmoothing()

        org_prompt = ""

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            negative_prompt,
            editing_prompt_embeds,
            negative_prompt_embeds,
            callback_on_step_end_tensor_inputs,
        )

        self._guidance_rescale = guidance_rescale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs

        # 2. Define call parameters
        batch_size = self.batch_size

        if editing_prompt:
            enable_edit_guidance = True
            if isinstance(editing_prompt, str):
                editing_prompt = [editing_prompt]
            self.enabled_editing_prompts = len(editing_prompt)
        elif editing_prompt_embeds is not None:
            enable_edit_guidance = True
            self.enabled_editing_prompts = editing_prompt_embeds.shape[0]
        else:
            self.enabled_editing_prompts = 0
            enable_edit_guidance = False

        # 3. Encode input prompt
        lora_scale = self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None

        edit_concepts, uncond_embeddings, num_edit_tokens = self.encode_prompt(
            editing_prompt=editing_prompt,
            num_images_per_prompt=num_images_per_prompt,
            enable_edit_guidance=enable_edit_guidance,
            negative_prompt=negative_prompt,
            editing_prompt_embeds=editing_prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            lora_scale=lora_scale,
            clip_skip=self.clip_skip,
        )

        # For classifier free guidance, we need to do two forward passes.
        # Here we concatenate the unconditional and text embeddings into a single batch
        # to avoid doing two forward passes
        if enable_edit_guidance:
            text_embeddings = mint.cat([uncond_embeddings, edit_concepts])
            self.text_cross_attention_maps = [editing_prompt] if isinstance(editing_prompt, str) else editing_prompt
        else:
            text_embeddings = mint.cat([uncond_embeddings])

        # 4. Prepare timesteps
        # self.scheduler.set_timesteps(num_inference_steps)
        timesteps = self.inversion_steps
        t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs.shape[0] :])}

        if use_cross_attn_mask:
            self.attention_store = LeditsAttentionStore(
                average=store_averaged_over_steps,
                batch_size=batch_size,
                max_size=(latents.shape[-2] / 4.0) * (latents.shape[-1] / 4.0),
                max_resolution=None,
            )
            self.prepare_unet(self.attention_store, PnP=False)
            resolution = latents.shape[-2:]
            att_res = (int(resolution[0] / 4), int(resolution[1] / 4))

        # 5. Prepare latent variables
        num_channels_latents = self.unet.config.in_channels
        latents = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            None,
            None,
            text_embeddings.dtype,
            latents,
        )

        # 6. Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(eta)

        self.sem_guidance = None
        self.activation_mask = None

        # 7. Denoising loop
        num_warmup_steps = 0
        with pynative_context():
            with self.progress_bar(total=len(timesteps)) as progress_bar:
                for i, t in enumerate(timesteps):
                    # expand the latents if we are doing classifier free guidance

                    if enable_edit_guidance:
                        latent_model_input = mint.cat([latents] * (1 + self.enabled_editing_prompts))
                    else:
                        latent_model_input = latents

                    latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                    text_embed_input = text_embeddings

                    # predict the noise residual
                    noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input)[0]

                    noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts)  # [b,4, 64, 64]
                    noise_pred_uncond = noise_pred_out[0]
                    noise_pred_edit_concepts = noise_pred_out[1:]

                    noise_guidance_edit = mint.zeros(
                        noise_pred_uncond.shape,
                        dtype=noise_pred_uncond.dtype,
                    )

                    if sem_guidance is not None and len(sem_guidance) > i:
                        noise_guidance_edit += sem_guidance[i]

                    elif enable_edit_guidance:
                        if self.activation_mask is None:
                            self.activation_mask = mint.zeros(
                                (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
                            )

                        if self.sem_guidance is None:
                            self.sem_guidance = mint.zeros((len(timesteps), *noise_pred_uncond.shape))

                        for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
                            if isinstance(edit_warmup_steps, list):
                                edit_warmup_steps_c = edit_warmup_steps[c]
                            else:
                                edit_warmup_steps_c = edit_warmup_steps
                            if i < edit_warmup_steps_c:
                                continue

                            if isinstance(edit_guidance_scale, list):
                                edit_guidance_scale_c = edit_guidance_scale[c]
                            else:
                                edit_guidance_scale_c = edit_guidance_scale

                            if isinstance(edit_threshold, list):
                                edit_threshold_c = edit_threshold[c]
                            else:
                                edit_threshold_c = edit_threshold
                            if isinstance(reverse_editing_direction, list):
                                reverse_editing_direction_c = reverse_editing_direction[c]
                            else:
                                reverse_editing_direction_c = reverse_editing_direction

                            if isinstance(edit_cooldown_steps, list):
                                edit_cooldown_steps_c = edit_cooldown_steps[c]
                            elif edit_cooldown_steps is None:
                                edit_cooldown_steps_c = i + 1
                            else:
                                edit_cooldown_steps_c = edit_cooldown_steps

                            if i >= edit_cooldown_steps_c:
                                continue

                            noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond

                            if reverse_editing_direction_c:
                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1

                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c

                            if user_mask is not None:
                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask

                            if use_cross_attn_mask:
                                out = self.attention_store.aggregate_attention(
                                    attention_maps=self.attention_store.step_store,
                                    prompts=self.text_cross_attention_maps,
                                    res=att_res,
                                    from_where=["up", "down"],
                                    is_cross=True,
                                    select=self.text_cross_attention_maps.index(editing_prompt[c]),
                                )
                                attn_map = out[:, :, :, 1 : 1 + num_edit_tokens[c].item()]  # 0 -> startoftext

                                # average over all tokens
                                if attn_map.shape[3] != num_edit_tokens[c]:
                                    raise ValueError(
                                        f"Incorrect shape of attention_map. Expected size {num_edit_tokens[c]}, but found {attn_map.shape[3]}!"
                                    )

                                attn_map = mint.sum(attn_map, dim=3)

                                # gaussian_smoothing
                                attn_map = mint.nn.functional.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect")
                                attn_map = self.smoothing(attn_map).squeeze(1)

                                # ops.quantile function expects float32
                                # TODO: ops.quantile is not supported
                                if attn_map.dtype == ms.float32:
                                    tmp = ms.tensor(
                                        np.quantile(attn_map.flatten(start_dim=1).numpy(), edit_threshold_c, axis=1)
                                    )
                                else:
                                    tmp = ms.tensor(
                                        np.quantile(
                                            attn_map.flatten(start_dim=1).to(ms.float32).numpy(),
                                            edit_threshold_c,
                                            axis=1,
                                        )
                                    ).to(attn_map.dtype)
                                attn_mask = mint.where(
                                    attn_map >= tmp.unsqueeze(1).unsqueeze(1).tile((1, *att_res)), 1.0, 0.0
                                )

                                # resolution must match latent space dimension
                                attn_mask = mint.nn.functional.interpolate(
                                    attn_mask.unsqueeze(1),
                                    noise_guidance_edit_tmp.shape[-2:],  # 64,64
                                ).tile((1, 4, 1, 1))
                                self.activation_mask[i, c] = ops.stop_gradient(attn_mask)
                                if not use_intersect_mask:
                                    noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                            if use_intersect_mask:
                                if t <= 800:
                                    noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                                    noise_guidance_edit_tmp_quantile = mint.sum(
                                        noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                                    )
                                    noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile(
                                        (1, self.unet.config.in_channels, 1, 1)
                                    )

                                    # ops.quantile function expects float32
                                    # TODO: ops.quantile is not supported
                                    if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                                        tmp = ms.tensor(
                                            np.quantile(
                                                noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                                edit_threshold_c,
                                                axis=2,
                                                keepdims=False,
                                            )
                                        )
                                    else:
                                        tmp = ms.tensor(
                                            np.quantile(
                                                noise_guidance_edit_tmp_quantile.flatten(start_dim=2)
                                                .to(ms.float32)
                                                .numpy(),
                                                edit_threshold_c,
                                                axis=2,
                                                keepdims=False,
                                            )
                                        ).to(noise_guidance_edit_tmp_quantile.dtype)

                                    intersect_mask = (
                                        mint.where(
                                            noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                            mint.ones_like(noise_guidance_edit_tmp),
                                            mint.zeros_like(noise_guidance_edit_tmp),
                                        )
                                        * attn_mask
                                    )

                                    self.activation_mask[i, c] = ops.stop_gradient(intersect_mask)

                                    noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask

                                else:
                                    # print(f"only attention mask for step {i}")
                                    noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                            elif not use_cross_attn_mask:
                                # calculate quantile
                                noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                                noise_guidance_edit_tmp_quantile = mint.sum(
                                    noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                                )
                                noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile((1, 4, 1, 1))

                                # ops.quantile function expects float32
                                # TODO: ops.quantile is not supported
                                if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                                    tmp = ms.tensor(
                                        np.quantile(
                                            noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                            edit_threshold_c,
                                            axis=2,
                                            keepdims=False,
                                        )
                                    )
                                else:
                                    tmp = ms.tensor(
                                        np.quantile(
                                            noise_guidance_edit_tmp_quantile.flatten(start_dim=2)
                                            .to(ms.float32)
                                            .numpy(),
                                            edit_threshold_c,
                                            axis=2,
                                            keepdims=False,
                                        )
                                    ).to(noise_guidance_edit_tmp_quantile.dtype)

                                self.activation_mask[i, c] = ops.stop_gradient(
                                    mint.where(
                                        noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                        mint.ones_like(noise_guidance_edit_tmp),
                                        mint.zeros_like(noise_guidance_edit_tmp),
                                    )
                                )

                                noise_guidance_edit_tmp = mint.where(
                                    noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                    noise_guidance_edit_tmp,
                                    mint.zeros_like(noise_guidance_edit_tmp),
                                )

                            noise_guidance_edit += noise_guidance_edit_tmp

                        self.sem_guidance[i] = ops.stop_gradient(noise_guidance_edit)

                    noise_pred = noise_pred_uncond + noise_guidance_edit

                    if enable_edit_guidance and self.guidance_rescale > 0.0:
                        # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                        noise_pred = rescale_noise_cfg(
                            noise_pred,
                            noise_pred_edit_concepts.mean(dim=0, keepdim=False),
                            guidance_rescale=self.guidance_rescale,
                        )

                    idx = t_to_idx[int(t)]
                    latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs)[
                        0
                    ]

                    # step callback
                    if use_cross_attn_mask:
                        store_step = i in attn_store_steps
                        self.attention_store.between_steps(store_step)

                    if callback_on_step_end is not None:
                        callback_kwargs = {}
                        for k in callback_on_step_end_tensor_inputs:
                            callback_kwargs[k] = locals()[k]
                        callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                        latents = callback_outputs.pop("latents", latents)
                        # prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                        negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)

                    # call the callback, if provided
                    if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                        progress_bar.update()

        # 8. Post-processing
        if not output_type == "latent":
            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
            image, has_nsfw_concept = self.run_safety_checker(image, text_embeddings.dtype)
        else:
            image = latents
            has_nsfw_concept = None

        if has_nsfw_concept is None:
            do_denormalize = [True] * image.shape[0]
        else:
            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

        if not return_dict:
            return (image, has_nsfw_concept)

        return LEditsPPDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

    def invert(
        self,
        image: PipelineImageInput,
        source_prompt: str = "",
        source_guidance_scale: float = 3.5,
        num_inversion_steps: int = 30,
        skip: float = 0.15,
        generator: Optional[np.random.Generator] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        clip_skip: Optional[int] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        resize_mode: Optional[str] = "default",
        crops_coords: Optional[Tuple[int, int, int, int]] = None,
    ):
        r"""
        The function to the pipeline for image inversion as described by the [LEDITS++
        Paper](https://arxiv.org/abs/2301.12247). If the scheduler is set to [`~schedulers.DDIMScheduler`] the
        inversion proposed by [edit-friendly DPDM](https://arxiv.org/abs/2304.06140) will be performed instead.

        Args:
            image (`PipelineImageInput`):
                Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect
                ratio.
            source_prompt (`str`, defaults to `""`):
                Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled
                if the `source_prompt` is `""`.
            source_guidance_scale (`float`, defaults to `3.5`):
                Strength of guidance during inversion.
            num_inversion_steps (`int`, defaults to `30`):
                Number of total performed inversion steps after discarding the initial `skip` steps.
            skip (`float`, defaults to `0.15`):
                Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values
                will lead to stronger changes to the input image. `skip` has to be between `0` and `1`.
            generator (`np.random.Generator`, *optional*):
                A [`np.random.Generator`](https://numpy.org/doc/stable/reference/random/generator.html) to make inversion
                deterministic.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            height (`int`, *optional*, defaults to `None`):
                The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default
                height.
            width (`int`, *optional*`, defaults to `None`):
                The width in preprocessed. If `None`, will use get_default_height_width()` to get the default width.
            resize_mode (`str`, *optional*, defaults to `default`):
                The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit within
                the specified width and height, and it may not maintaining the original aspect ratio. If `fill`, will
                resize the image to fit within the specified width and height, maintaining the aspect ratio, and then
                center the image within the dimensions, filling empty with data from image. If `crop`, will resize the
                image to fit within the specified width and height, maintaining the aspect ratio, and then center the
                image within the dimensions, cropping the excess. Note that resize_mode `fill` and `crop` are only
                supported for PIL image input.
            crops_coords (`List[Tuple[int, int, int, int]]`, *optional*, defaults to `None`):
                The crop coordinates for each image in the batch. If `None`, will not crop the image.

        Returns:
            [`~pipelines.ledits_pp.LEditsPPInversionPipelineOutput`]: Output will contain the resized input image(s)
            and respective VAE reconstruction(s).
        """
        if height is not None and height % 32 != 0 or width is not None and width % 32 != 0:
            raise ValueError("height and width must be a factor of 32.")
        # Reset attn processor, we do not want to store attn maps during inversion
        self.unet.set_attn_processor(AttnProcessor())

        self.eta = 1.0

        self.scheduler.config.timestep_spacing = "leading"
        self.scheduler.set_timesteps(int(num_inversion_steps * (1 + skip)))
        self.inversion_steps = self.scheduler.timesteps[-num_inversion_steps:]
        timesteps = self.inversion_steps

        # 1. encode image
        x0, resized = self.encode_image(
            image,
            dtype=self.text_encoder.dtype,
            height=height,
            width=width,
            resize_mode=resize_mode,
            crops_coords=crops_coords,
        )
        self.batch_size = x0.shape[0]

        # autoencoder reconstruction
        image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
        image_rec = self.image_processor.postprocess(image_rec, output_type="pil")

        # 2. get embeddings
        do_classifier_free_guidance = source_guidance_scale > 1.0

        lora_scale = cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None

        uncond_embedding, text_embeddings, _ = self.encode_prompt(
            num_images_per_prompt=1,
            negative_prompt=None,
            enable_edit_guidance=do_classifier_free_guidance,
            editing_prompt=source_prompt,
            lora_scale=lora_scale,
            clip_skip=clip_skip,
        )

        # 3. find zs and xts
        variance_noise_shape = (num_inversion_steps, *x0.shape)

        # intermediate latents
        t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
        xts = mint.zeros(size=variance_noise_shape, dtype=uncond_embedding.dtype)

        for t in reversed(timesteps):
            idx = num_inversion_steps - t_to_idx[int(t)] - 1
            noise = randn_tensor(shape=x0.shape, generator=generator, dtype=x0.dtype)
            xts[idx] = self.scheduler.add_noise(x0, noise, ms.tensor([t]))
        xts = mint.cat([x0.unsqueeze(0), xts], dim=0)

        self.scheduler.set_timesteps(len(self.scheduler.timesteps))
        # noise maps
        zs = mint.zeros(size=variance_noise_shape, dtype=uncond_embedding.dtype)

        with self.progress_bar(total=len(timesteps)) as progress_bar:
            for t in timesteps:
                idx = num_inversion_steps - t_to_idx[int(t)] - 1
                # 1. predict noise residual
                xt = xts[idx + 1]

                noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=uncond_embedding)[0]

                if not source_prompt == "":
                    noise_pred_cond = self.unet(xt, timestep=t, encoder_hidden_states=text_embeddings)[0]
                    noise_pred = noise_pred + source_guidance_scale * (noise_pred_cond - noise_pred)

                xtm1 = xts[idx]
                z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, self.eta)
                zs[idx] = z

                # correction to avoid error accumulation
                xts[idx] = xtm1_corrected

                progress_bar.update()

        self.init_latents = xts[-1].broadcast_to((self.batch_size, -1, -1, -1))
        zs = zs.flip((0,))
        self.zs = zs

        return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec)

    def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="default", crops_coords=None):
        image = self.image_processor.preprocess(
            image=image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
        )
        height, width = image.shape[-2:]
        if height % 32 != 0 or width % 32 != 0:
            raise ValueError(
                "Image height and width must be a factor of 32. "
                "Consider down-sampling the input using the `height` and `width` parameters"
            )
        resized = self.image_processor.postprocess(image=image, output_type="pil")

        if max(image.shape[-2:]) > self.vae.config["sample_size"] * 1.5:
            logger.warning(
                "Your input images far exceed the default resolution of the underlying diffusion model. "
                "The output images may contain severe artifacts! "
                "Consider down-sampling the input using the `height` and `width` parameters"
            )
        image = image.to(dtype)

        x0 = self.vae.diag_gauss_dist.mode(self.vae.encode(image)[0])
        x0 = x0.to(dtype)
        x0 = self.vae.config.scaling_factor * x0
        return x0, resized

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.__call__(negative_prompt=None, generator=None, output_type='pil', return_dict=False, editing_prompt=None, editing_prompt_embeds=None, negative_prompt_embeds=None, reverse_editing_direction=False, edit_guidance_scale=5, edit_warmup_steps=0, edit_cooldown_steps=None, edit_threshold=0.9, user_mask=None, sem_guidance=None, use_cross_attn_mask=False, use_intersect_mask=True, attn_store_steps=[], store_averaged_over_steps=True, cross_attention_kwargs=None, guidance_rescale=0.0, clip_skip=None, callback_on_step_end=None, callback_on_step_end_tensor_inputs=['latents'], **kwargs)

The call function to the pipeline for editing. The [~pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.invert] method has to be called beforehand. Edits will always be performed for the last inverted image(s).

PARAMETER	DESCRIPTION
negative_prompt	The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored if guidance_scale is less than 1). TYPE: `str` or `List[str]`, *optional* DEFAULT: None
generator	One or a list of np.random.Generator(s) to make generation deterministic. TYPE: `np.random.Generator`, *optional* DEFAULT: None
output_type	The output format of the generate image. Choose between PIL: PIL.Image.Image or np.array. TYPE: `str`, *optional*, defaults to `"pil"` DEFAULT: 'pil'
return_dict	Whether or not to return a [~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput] instead of a plain tuple. TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False
editing_prompt	The prompt or prompts to guide the image generation. The image is reconstructed by setting editing_prompt = None. Guidance direction of prompt should be specified via reverse_editing_direction. TYPE: `str` or `List[str]`, *optional* DEFAULT: None
editing_prompt_embeds	Pre-computed embeddings to use for guiding the image generation. Guidance direction of embedding should be specified via reverse_editing_direction. TYPE: `ms.Tensor>`, *optional* DEFAULT: None
negative_prompt_embeds	Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, negative_prompt_embeds are generated from the negative_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
reverse_editing_direction	Whether the corresponding prompt in editing_prompt should be increased or decreased. TYPE: `bool` or `List[bool]`, *optional*, defaults to `False` DEFAULT: False
edit_guidance_scale	Guidance scale for guiding the image generation. If provided as list values should correspond to editing_prompt. edit_guidance_scale is defined as s_e of equation 12 of LEDITS++ Paper. TYPE: `float` or `List[float]`, *optional*, defaults to 5 DEFAULT: 5
edit_warmup_steps	Number of diffusion steps (for each prompt) for which guidance will not be applied. TYPE: `float` or `List[float]`, *optional*, defaults to 10 DEFAULT: 0
edit_cooldown_steps	Number of diffusion steps (for each prompt) after which guidance will no longer be applied. TYPE: `float` or `List[float]`, *optional*, defaults to `None` DEFAULT: None
edit_threshold	Masking threshold of guidance. Threshold should be proportional to the image region that is modified. 'edit_threshold' is defined as 'λ' of equation 12 of LEDITS++ Paper. TYPE: `float` or `List[float]`, *optional*, defaults to 0.9 DEFAULT: 0.9
user_mask	User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s implicit masks do not meet user preferences. TYPE: `ms.Tensor`, *optional* DEFAULT: None
sem_guidance	List of pre-generated guidance vectors to be applied at generation. Length of the list has to correspond to num_inference_steps. TYPE: `List[ms.Tensor]`, *optional* DEFAULT: None
use_cross_attn_mask	Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of LEDITS++ paper. TYPE: `bool`, defaults to `False` DEFAULT: False
use_intersect_mask	Whether the masking term is calculated as intersection of cross-attention masks and masks derived from the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate are defined as 'M^2' of equation 12 of LEDITS++ paper. TYPE: `bool`, defaults to `True` DEFAULT: True
attn_store_steps	Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes. TYPE: `List[int]`, *optional* DEFAULT: []
store_averaged_over_steps	Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If False, attention maps for each step are stores separately. Just for visualization purposes. TYPE: `bool`, defaults to `True` DEFAULT: True
cross_attention_kwargs	A kwargs dictionary that if specified is passed along to the [AttentionProcessor] as defined in self.processor. TYPE: `dict`, *optional* DEFAULT: None
guidance_rescale	Guidance rescale factor from Common Diffusion Noise Schedules and Sample Steps are Flawed. Guidance rescale factor should fix overexposure when using zero terminal SNR. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
clip_skip	Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. TYPE: `int`, *optional* DEFAULT: None
callback_on_step_end	A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict). callback_kwargs will include a list of all tensors as specified by callback_on_step_end_tensor_inputs. TYPE: `Callable`, *optional* DEFAULT: None
callback_on_step_end_tensor_inputs	The list of tensor inputs for the callback_on_step_end function. The tensors specified in the list will be passed as callback_kwargs argument. You will only be able to include variables listed in the ._callback_tensor_inputs attribute of your pipeline class. TYPE: `List`, *optional* DEFAULT: ['latents']

RETURNS	DESCRIPTION
	[~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput] or tuple:
	[~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput] if return_dict is True, otherwise a `tuple. When
	returning a tuple, the first element is a list with the generated images, and the second element is a list
	of bools denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
	content, according to the safety_checker.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def __call__(
    self,
    negative_prompt: Optional[Union[str, List[str]]] = None,
    generator: Optional[Union[np.random.Generator, List[np.random.Generator]]] = None,
    output_type: Optional[str] = "pil",
    return_dict: bool = False,
    editing_prompt: Optional[Union[str, List[str]]] = None,
    editing_prompt_embeds: Optional[ms.Tensor] = None,
    negative_prompt_embeds: Optional[ms.Tensor] = None,
    reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
    edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
    edit_warmup_steps: Optional[Union[int, List[int]]] = 0,
    edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
    edit_threshold: Optional[Union[float, List[float]]] = 0.9,
    user_mask: Optional[ms.Tensor] = None,
    sem_guidance: Optional[List[ms.Tensor]] = None,
    use_cross_attn_mask: bool = False,
    use_intersect_mask: bool = True,
    attn_store_steps: Optional[List[int]] = [],
    store_averaged_over_steps: bool = True,
    cross_attention_kwargs: Optional[Dict[str, Any]] = None,
    guidance_rescale: float = 0.0,
    clip_skip: Optional[int] = None,
    callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
    callback_on_step_end_tensor_inputs: List[str] = ["latents"],
    **kwargs,
):
    r"""
    The call function to the pipeline for editing. The
    [`~pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.invert`] method has to be called beforehand. Edits will
    always be performed for the last inverted image(s).

    Args:
        negative_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation. Ignored when not using guidance (i.e., ignored
            if `guidance_scale` is less than `1`).
        generator (`np.random.Generator`, *optional*):
            One or a list of [np.random.Generator(s)](https://numpy.org/doc/stable/reference/random/generator.html)
            to make generation deterministic.
        output_type (`str`, *optional*, defaults to `"pil"`):
            The output format of the generate image. Choose between
            [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
        return_dict (`bool`, *optional*, defaults to `False`):
            Whether or not to return a [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] instead of a plain
            tuple.
        editing_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts to guide the image generation. The image is reconstructed by setting
            `editing_prompt = None`. Guidance direction of prompt should be specified via
            `reverse_editing_direction`.
        editing_prompt_embeds (`ms.Tensor>`, *optional*):
            Pre-computed embeddings to use for guiding the image generation. Guidance direction of embedding should
            be specified via `reverse_editing_direction`.
        negative_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
            not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
        reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
            Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
        edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
            Guidance scale for guiding the image generation. If provided as list values should correspond to
            `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 12 of [LEDITS++
            Paper](https://arxiv.org/abs/2301.12247).
        edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
            Number of diffusion steps (for each prompt) for which guidance will not be applied.
        edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
            Number of diffusion steps (for each prompt) after which guidance will no longer be applied.
        edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
            Masking threshold of guidance. Threshold should be proportional to the image region that is modified.
            'edit_threshold' is defined as 'λ' of equation 12 of [LEDITS++
            Paper](https://arxiv.org/abs/2301.12247).
        user_mask (`ms.Tensor`, *optional*):
            User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s
            implicit masks do not meet user preferences.
        sem_guidance (`List[ms.Tensor]`, *optional*):
            List of pre-generated guidance vectors to be applied at generation. Length of the list has to
            correspond to `num_inference_steps`.
        use_cross_attn_mask (`bool`, defaults to `False`):
            Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask
            is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of [LEDITS++
            paper](https://arxiv.org/pdf/2311.16711.pdf).
        use_intersect_mask (`bool`, defaults to `True`):
            Whether the masking term is calculated as intersection of cross-attention masks and masks derived from
            the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate
            are defined as 'M^2' of equation 12 of [LEDITS++ paper](https://arxiv.org/pdf/2311.16711.pdf).
        attn_store_steps (`List[int]`, *optional*):
            Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes.
        store_averaged_over_steps (`bool`, defaults to `True`):
            Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If
            False, attention maps for each step are stores separately. Just for visualization purposes.
        cross_attention_kwargs (`dict`, *optional*):
            A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
            [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
        guidance_rescale (`float`, *optional*, defaults to 0.0):
            Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
            Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
            using zero terminal SNR.
        clip_skip (`int`, *optional*):
            Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
            the output of the pre-final layer will be used for computing the prompt embeddings.
        callback_on_step_end (`Callable`, *optional*):
            A function that calls at the end of each denoising steps during the inference. The function is called
            with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
            callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
            `callback_on_step_end_tensor_inputs`.
        callback_on_step_end_tensor_inputs (`List`, *optional*):
            The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
            will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
            `._callback_tensor_inputs` attribute of your pipeline class.

    Examples:

    Returns:
        [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] or `tuple`:
        [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
        returning a tuple, the first element is a list with the generated images, and the second element is a list
        of `bool`s denoting whether the corresponding generated image likely represents "not-safe-for-work" (nsfw)
        content, according to the `safety_checker`.
    """

    if self.inversion_steps is None:
        raise ValueError(
            "You need to invert an input image first before calling the pipeline. The `invert` method has to be called beforehand. Edits will always be performed for the last inverted image(s)."  # noqa: E501
        )

    eta = self.eta
    num_images_per_prompt = 1
    latents = self.init_latents

    zs = self.zs
    self.scheduler.set_timesteps(len(self.scheduler.timesteps))

    if use_intersect_mask:
        use_cross_attn_mask = True

    if use_cross_attn_mask:
        self.smoothing = LeditsGaussianSmoothing()

    org_prompt = ""

    # 1. Check inputs. Raise error if not correct
    self.check_inputs(
        negative_prompt,
        editing_prompt_embeds,
        negative_prompt_embeds,
        callback_on_step_end_tensor_inputs,
    )

    self._guidance_rescale = guidance_rescale
    self._clip_skip = clip_skip
    self._cross_attention_kwargs = cross_attention_kwargs

    # 2. Define call parameters
    batch_size = self.batch_size

    if editing_prompt:
        enable_edit_guidance = True
        if isinstance(editing_prompt, str):
            editing_prompt = [editing_prompt]
        self.enabled_editing_prompts = len(editing_prompt)
    elif editing_prompt_embeds is not None:
        enable_edit_guidance = True
        self.enabled_editing_prompts = editing_prompt_embeds.shape[0]
    else:
        self.enabled_editing_prompts = 0
        enable_edit_guidance = False

    # 3. Encode input prompt
    lora_scale = self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None

    edit_concepts, uncond_embeddings, num_edit_tokens = self.encode_prompt(
        editing_prompt=editing_prompt,
        num_images_per_prompt=num_images_per_prompt,
        enable_edit_guidance=enable_edit_guidance,
        negative_prompt=negative_prompt,
        editing_prompt_embeds=editing_prompt_embeds,
        negative_prompt_embeds=negative_prompt_embeds,
        lora_scale=lora_scale,
        clip_skip=self.clip_skip,
    )

    # For classifier free guidance, we need to do two forward passes.
    # Here we concatenate the unconditional and text embeddings into a single batch
    # to avoid doing two forward passes
    if enable_edit_guidance:
        text_embeddings = mint.cat([uncond_embeddings, edit_concepts])
        self.text_cross_attention_maps = [editing_prompt] if isinstance(editing_prompt, str) else editing_prompt
    else:
        text_embeddings = mint.cat([uncond_embeddings])

    # 4. Prepare timesteps
    # self.scheduler.set_timesteps(num_inference_steps)
    timesteps = self.inversion_steps
    t_to_idx = {int(v): k for k, v in enumerate(timesteps[-zs.shape[0] :])}

    if use_cross_attn_mask:
        self.attention_store = LeditsAttentionStore(
            average=store_averaged_over_steps,
            batch_size=batch_size,
            max_size=(latents.shape[-2] / 4.0) * (latents.shape[-1] / 4.0),
            max_resolution=None,
        )
        self.prepare_unet(self.attention_store, PnP=False)
        resolution = latents.shape[-2:]
        att_res = (int(resolution[0] / 4), int(resolution[1] / 4))

    # 5. Prepare latent variables
    num_channels_latents = self.unet.config.in_channels
    latents = self.prepare_latents(
        batch_size * num_images_per_prompt,
        num_channels_latents,
        None,
        None,
        text_embeddings.dtype,
        latents,
    )

    # 6. Prepare extra step kwargs.
    extra_step_kwargs = self.prepare_extra_step_kwargs(eta)

    self.sem_guidance = None
    self.activation_mask = None

    # 7. Denoising loop
    num_warmup_steps = 0
    with pynative_context():
        with self.progress_bar(total=len(timesteps)) as progress_bar:
            for i, t in enumerate(timesteps):
                # expand the latents if we are doing classifier free guidance

                if enable_edit_guidance:
                    latent_model_input = mint.cat([latents] * (1 + self.enabled_editing_prompts))
                else:
                    latent_model_input = latents

                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

                text_embed_input = text_embeddings

                # predict the noise residual
                noise_pred = self.unet(latent_model_input, t, encoder_hidden_states=text_embed_input)[0]

                noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts)  # [b,4, 64, 64]
                noise_pred_uncond = noise_pred_out[0]
                noise_pred_edit_concepts = noise_pred_out[1:]

                noise_guidance_edit = mint.zeros(
                    noise_pred_uncond.shape,
                    dtype=noise_pred_uncond.dtype,
                )

                if sem_guidance is not None and len(sem_guidance) > i:
                    noise_guidance_edit += sem_guidance[i]

                elif enable_edit_guidance:
                    if self.activation_mask is None:
                        self.activation_mask = mint.zeros(
                            (len(timesteps), len(noise_pred_edit_concepts), *noise_pred_edit_concepts[0].shape)
                        )

                    if self.sem_guidance is None:
                        self.sem_guidance = mint.zeros((len(timesteps), *noise_pred_uncond.shape))

                    for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
                        if isinstance(edit_warmup_steps, list):
                            edit_warmup_steps_c = edit_warmup_steps[c]
                        else:
                            edit_warmup_steps_c = edit_warmup_steps
                        if i < edit_warmup_steps_c:
                            continue

                        if isinstance(edit_guidance_scale, list):
                            edit_guidance_scale_c = edit_guidance_scale[c]
                        else:
                            edit_guidance_scale_c = edit_guidance_scale

                        if isinstance(edit_threshold, list):
                            edit_threshold_c = edit_threshold[c]
                        else:
                            edit_threshold_c = edit_threshold
                        if isinstance(reverse_editing_direction, list):
                            reverse_editing_direction_c = reverse_editing_direction[c]
                        else:
                            reverse_editing_direction_c = reverse_editing_direction

                        if isinstance(edit_cooldown_steps, list):
                            edit_cooldown_steps_c = edit_cooldown_steps[c]
                        elif edit_cooldown_steps is None:
                            edit_cooldown_steps_c = i + 1
                        else:
                            edit_cooldown_steps_c = edit_cooldown_steps

                        if i >= edit_cooldown_steps_c:
                            continue

                        noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond

                        if reverse_editing_direction_c:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1

                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c

                        if user_mask is not None:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask

                        if use_cross_attn_mask:
                            out = self.attention_store.aggregate_attention(
                                attention_maps=self.attention_store.step_store,
                                prompts=self.text_cross_attention_maps,
                                res=att_res,
                                from_where=["up", "down"],
                                is_cross=True,
                                select=self.text_cross_attention_maps.index(editing_prompt[c]),
                            )
                            attn_map = out[:, :, :, 1 : 1 + num_edit_tokens[c].item()]  # 0 -> startoftext

                            # average over all tokens
                            if attn_map.shape[3] != num_edit_tokens[c]:
                                raise ValueError(
                                    f"Incorrect shape of attention_map. Expected size {num_edit_tokens[c]}, but found {attn_map.shape[3]}!"
                                )

                            attn_map = mint.sum(attn_map, dim=3)

                            # gaussian_smoothing
                            attn_map = mint.nn.functional.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect")
                            attn_map = self.smoothing(attn_map).squeeze(1)

                            # ops.quantile function expects float32
                            # TODO: ops.quantile is not supported
                            if attn_map.dtype == ms.float32:
                                tmp = ms.tensor(
                                    np.quantile(attn_map.flatten(start_dim=1).numpy(), edit_threshold_c, axis=1)
                                )
                            else:
                                tmp = ms.tensor(
                                    np.quantile(
                                        attn_map.flatten(start_dim=1).to(ms.float32).numpy(),
                                        edit_threshold_c,
                                        axis=1,
                                    )
                                ).to(attn_map.dtype)
                            attn_mask = mint.where(
                                attn_map >= tmp.unsqueeze(1).unsqueeze(1).tile((1, *att_res)), 1.0, 0.0
                            )

                            # resolution must match latent space dimension
                            attn_mask = mint.nn.functional.interpolate(
                                attn_mask.unsqueeze(1),
                                noise_guidance_edit_tmp.shape[-2:],  # 64,64
                            ).tile((1, 4, 1, 1))
                            self.activation_mask[i, c] = ops.stop_gradient(attn_mask)
                            if not use_intersect_mask:
                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                        if use_intersect_mask:
                            if t <= 800:
                                noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                                noise_guidance_edit_tmp_quantile = mint.sum(
                                    noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                                )
                                noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile(
                                    (1, self.unet.config.in_channels, 1, 1)
                                )

                                # ops.quantile function expects float32
                                # TODO: ops.quantile is not supported
                                if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                                    tmp = ms.tensor(
                                        np.quantile(
                                            noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                            edit_threshold_c,
                                            axis=2,
                                            keepdims=False,
                                        )
                                    )
                                else:
                                    tmp = ms.tensor(
                                        np.quantile(
                                            noise_guidance_edit_tmp_quantile.flatten(start_dim=2)
                                            .to(ms.float32)
                                            .numpy(),
                                            edit_threshold_c,
                                            axis=2,
                                            keepdims=False,
                                        )
                                    ).to(noise_guidance_edit_tmp_quantile.dtype)

                                intersect_mask = (
                                    mint.where(
                                        noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                        mint.ones_like(noise_guidance_edit_tmp),
                                        mint.zeros_like(noise_guidance_edit_tmp),
                                    )
                                    * attn_mask
                                )

                                self.activation_mask[i, c] = ops.stop_gradient(intersect_mask)

                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask

                            else:
                                # print(f"only attention mask for step {i}")
                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                        elif not use_cross_attn_mask:
                            # calculate quantile
                            noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                            noise_guidance_edit_tmp_quantile = mint.sum(
                                noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                            )
                            noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile((1, 4, 1, 1))

                            # ops.quantile function expects float32
                            # TODO: ops.quantile is not supported
                            if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                                tmp = ms.tensor(
                                    np.quantile(
                                        noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                        edit_threshold_c,
                                        axis=2,
                                        keepdims=False,
                                    )
                                )
                            else:
                                tmp = ms.tensor(
                                    np.quantile(
                                        noise_guidance_edit_tmp_quantile.flatten(start_dim=2)
                                        .to(ms.float32)
                                        .numpy(),
                                        edit_threshold_c,
                                        axis=2,
                                        keepdims=False,
                                    )
                                ).to(noise_guidance_edit_tmp_quantile.dtype)

                            self.activation_mask[i, c] = ops.stop_gradient(
                                mint.where(
                                    noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                    mint.ones_like(noise_guidance_edit_tmp),
                                    mint.zeros_like(noise_guidance_edit_tmp),
                                )
                            )

                            noise_guidance_edit_tmp = mint.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                noise_guidance_edit_tmp,
                                mint.zeros_like(noise_guidance_edit_tmp),
                            )

                        noise_guidance_edit += noise_guidance_edit_tmp

                    self.sem_guidance[i] = ops.stop_gradient(noise_guidance_edit)

                noise_pred = noise_pred_uncond + noise_guidance_edit

                if enable_edit_guidance and self.guidance_rescale > 0.0:
                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                    noise_pred = rescale_noise_cfg(
                        noise_pred,
                        noise_pred_edit_concepts.mean(dim=0, keepdim=False),
                        guidance_rescale=self.guidance_rescale,
                    )

                idx = t_to_idx[int(t)]
                latents = self.scheduler.step(noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs)[
                    0
                ]

                # step callback
                if use_cross_attn_mask:
                    store_step = i in attn_store_steps
                    self.attention_store.between_steps(store_step)

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    # prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

    # 8. Post-processing
    if not output_type == "latent":
        image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
        image, has_nsfw_concept = self.run_safety_checker(image, text_embeddings.dtype)
    else:
        image = latents
        has_nsfw_concept = None

    if has_nsfw_concept is None:
        do_denormalize = [True] * image.shape[0]
    else:
        do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]

    image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)

    if not return_dict:
        return (image, has_nsfw_concept)

    return LEditsPPDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.disable_vae_slicing()

Disable sliced VAE decoding. If enable_vae_slicing was previously enabled, this method will go back to computing decoding in one step.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def disable_vae_slicing(self):
    r"""
    Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
    computing decoding in one step.
    """
    self.vae.disable_slicing()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.disable_vae_tiling()

Disable tiled VAE decoding. If enable_vae_tiling was previously enabled, this method will go back to computing decoding in one step.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def disable_vae_tiling(self):
    r"""
    Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
    computing decoding in one step.
    """
    self.vae.disable_tiling()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.enable_vae_slicing()

Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def enable_vae_slicing(self):
    r"""
    Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
    compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
    """
    self.vae.enable_slicing()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.enable_vae_tiling()

Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow processing larger images.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def enable_vae_tiling(self):
    r"""
    Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
    compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
    processing larger images.
    """
    self.vae.enable_tiling()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.encode_prompt(num_images_per_prompt, enable_edit_guidance, negative_prompt=None, editing_prompt=None, negative_prompt_embeds=None, editing_prompt_embeds=None, lora_scale=None, clip_skip=None)

Encodes the prompt into text encoder hidden states.

PARAMETER	DESCRIPTION
num_images_per_prompt	number of images that should be generated per prompt TYPE: `int`
enable_edit_guidance	whether to perform any editing or reconstruct the input image instead TYPE: `bool`
negative_prompt	The prompt or prompts not to guide the image generation. If not defined, one has to pass negative_prompt_embeds instead. Ignored when not using guidance (i.e., ignored if guidance_scale is less than 1). TYPE: `str` or `List[str]`, *optional* DEFAULT: None
editing_prompt	Editing prompt(s) to be encoded. If not defined, one has to pass editing_prompt_embeds instead. TYPE: `str` or `List[str]`, *optional* DEFAULT: None
editing_prompt_embeds	Pre-generated text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, text embeddings will be generated from prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
negative_prompt_embeds	Pre-generated negative text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, negative_prompt_embeds will be generated from negative_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
lora_scale	A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. TYPE: `float`, *optional* DEFAULT: None
clip_skip	Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. TYPE: `int`, *optional* DEFAULT: None

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def encode_prompt(
    self,
    num_images_per_prompt,
    enable_edit_guidance,
    negative_prompt=None,
    editing_prompt=None,
    negative_prompt_embeds: Optional[ms.Tensor] = None,
    editing_prompt_embeds: Optional[ms.Tensor] = None,
    lora_scale: Optional[float] = None,
    clip_skip: Optional[int] = None,
):
    r"""
    Encodes the prompt into text encoder hidden states.

    Args:
        num_images_per_prompt (`int`):
            number of images that should be generated per prompt
        enable_edit_guidance (`bool`):
            whether to perform any editing or reconstruct the input image instead
        negative_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation. If not defined, one has to pass
            `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
            less than `1`).
        editing_prompt (`str` or `List[str]`, *optional*):
            Editing prompt(s) to be encoded. If not defined, one has to pass `editing_prompt_embeds` instead.
        editing_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
            provided, text embeddings will be generated from `prompt` input argument.
        negative_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
            argument.
        lora_scale (`float`, *optional*):
            A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        clip_skip (`int`, *optional*):
            Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
            the output of the pre-final layer will be used for computing the prompt embeddings.
    """
    # set lora scale so that monkey patched LoRA
    # function of text encoder can correctly access it
    if lora_scale is not None and isinstance(self, StableDiffusionLoraLoaderMixin):
        self._lora_scale = lora_scale

        # dynamically adjust the LoRA scale
        scale_lora_layers(self.text_encoder, lora_scale)

    batch_size = self.batch_size
    num_edit_tokens = None

    if negative_prompt_embeds is None:
        uncond_tokens: List[str]
        if negative_prompt is None:
            uncond_tokens = [""] * batch_size
        elif isinstance(negative_prompt, str):
            uncond_tokens = [negative_prompt]
        elif batch_size != len(negative_prompt):
            raise ValueError(
                f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but exoected"
                f"{batch_size} based on the input images. Please make sure that passed `negative_prompt` matches"
                " the batch size of `prompt`."
            )
        else:
            uncond_tokens = negative_prompt

        # textual inversion: procecss multi-vector tokens if necessary
        if isinstance(self, TextualInversionLoaderMixin):
            uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)

        uncond_input = self.tokenizer(
            uncond_tokens,
            padding="max_length",
            max_length=self.tokenizer.model_max_length,
            truncation=True,
            return_tensors="np",
        )

        if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
            attention_mask = ms.tensor(uncond_input.attention_mask)
        else:
            attention_mask = None

        negative_prompt_embeds = self.text_encoder(
            ms.tensor(uncond_input.input_ids),
            attention_mask=attention_mask,
        )
        negative_prompt_embeds = negative_prompt_embeds[0]

    if self.text_encoder is not None:
        prompt_embeds_dtype = self.text_encoder.dtype
    elif self.unet is not None:
        prompt_embeds_dtype = self.unet.dtype
    else:
        prompt_embeds_dtype = negative_prompt_embeds.dtype

    negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype)

    if enable_edit_guidance:
        if editing_prompt_embeds is None:
            # textual inversion: procecss multi-vector tokens if necessary
            # if isinstance(self, TextualInversionLoaderMixin):
            #    prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
            if isinstance(editing_prompt, str):
                editing_prompt = [editing_prompt]

            max_length = negative_prompt_embeds.shape[1]
            text_inputs = self.tokenizer(
                [x for item in editing_prompt for x in repeat(item, batch_size)],
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="np",
                return_length=True,
            )

            num_edit_tokens = text_inputs.length - 2  # not counting startoftext and endoftext
            text_input_ids = text_inputs.input_ids
            untruncated_ids = self.tokenizer(
                [x for item in editing_prompt for x in repeat(item, batch_size)],
                padding="longest",
                return_tensors="np",
            ).input_ids

            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not np.array_equal(
                text_input_ids, untruncated_ids
            ):
                removed_text = self.tokenizer.batch_decode(
                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
                )
                logger.warning(
                    "The following part of your input was truncated because CLIP can only handle sequences up to"
                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
                )

            if (
                hasattr(self.text_encoder.config, "use_attention_mask")
                and self.text_encoder.config.use_attention_mask
            ):
                attention_mask = ms.tensor(text_inputs.attention_mask)
            else:
                attention_mask = None

            text_input_ids = ms.tensor(text_input_ids)
            if clip_skip is None:
                editing_prompt_embeds = self.text_encoder(text_input_ids, attention_mask=attention_mask)
                editing_prompt_embeds = editing_prompt_embeds[0]
            else:
                editing_prompt_embeds = self.text_encoder(
                    text_input_ids, attention_mask=attention_mask, output_hidden_states=True
                )
                # Access the `hidden_states` first, that contains a tuple of
                # all the hidden states from the encoder layers. Then index into
                # the tuple to access the hidden states from the desired layer.
                editing_prompt_embeds = editing_prompt_embeds[-1][-(clip_skip + 1)]
                # We also need to apply the final LayerNorm here to not mess with the
                # representations. The `last_hidden_states` that we typically use for
                # obtaining the final prompt representations passes through the LayerNorm
                # layer.
                editing_prompt_embeds = self.text_encoder.text_model.final_layer_norm(editing_prompt_embeds)

        editing_prompt_embeds = editing_prompt_embeds.to(dtype=negative_prompt_embeds.dtype)

        bs_embed_edit, seq_len, _ = editing_prompt_embeds.shape
        editing_prompt_embeds = editing_prompt_embeds.to(dtype=negative_prompt_embeds.dtype)
        editing_prompt_embeds = editing_prompt_embeds.tile((1, num_images_per_prompt, 1))
        editing_prompt_embeds = editing_prompt_embeds.view(bs_embed_edit * num_images_per_prompt, seq_len, -1)

    # get unconditional embeddings for classifier free guidance

    # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
    seq_len = negative_prompt_embeds.shape[1]

    negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype)

    negative_prompt_embeds = negative_prompt_embeds.tile((1, num_images_per_prompt, 1))
    negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

    if isinstance(self, StableDiffusionLoraLoaderMixin):
        # Retrieve the original scale by scaling back the LoRA layers
        unscale_lora_layers(self.text_encoder, lora_scale)

    return editing_prompt_embeds, negative_prompt_embeds, num_edit_tokens

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusion.invert(image, source_prompt='', source_guidance_scale=3.5, num_inversion_steps=30, skip=0.15, generator=None, cross_attention_kwargs=None, clip_skip=None, height=None, width=None, resize_mode='default', crops_coords=None)

The function to the pipeline for image inversion as described by the LEDITS++ Paper. If the scheduler is set to [~schedulers.DDIMScheduler] the inversion proposed by edit-friendly DPDM will be performed instead.

PARAMETER	DESCRIPTION
image	Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect ratio. TYPE: `PipelineImageInput`
source_prompt	Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled if the source_prompt is "". TYPE: `str`, defaults to `""` DEFAULT: ''
source_guidance_scale	Strength of guidance during inversion. TYPE: `float`, defaults to `3.5` DEFAULT: 3.5
num_inversion_steps	Number of total performed inversion steps after discarding the initial skip steps. TYPE: `int`, defaults to `30` DEFAULT: 30
skip	Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values will lead to stronger changes to the input image. skip has to be between 0 and 1. TYPE: `float`, defaults to `0.15` DEFAULT: 0.15
generator	A np.random.Generator to make inversion deterministic. TYPE: `np.random.Generator`, *optional* DEFAULT: None
cross_attention_kwargs	A kwargs dictionary that if specified is passed along to the [AttentionProcessor] as defined in self.processor. TYPE: `dict`, *optional* DEFAULT: None
clip_skip	Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. TYPE: `int`, *optional* DEFAULT: None
height	The height in preprocessed image. If None, will use the get_default_height_width() to get default height. TYPE: `int`, *optional*, defaults to `None` DEFAULT: None
width	The width in preprocessed. If None, will use get_default_height_width()` to get the default width. TYPE: `int`, *optional*`, defaults to `None` DEFAULT: None
resize_mode	The resize mode, can be one of default or fill. If default, will resize the image to fit within the specified width and height, and it may not maintaining the original aspect ratio. If fill, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image within the dimensions, filling empty with data from image. If crop, will resize the image to fit within the specified width and height, maintaining the aspect ratio, and then center the image within the dimensions, cropping the excess. Note that resize_mode fill and crop are only supported for PIL image input. TYPE: `str`, *optional*, defaults to `default` DEFAULT: 'default'
crops_coords	The crop coordinates for each image in the batch. If None, will not crop the image. TYPE: `List[Tuple[int, int, int, int]]`, *optional*, defaults to `None` DEFAULT: None

RETURNS	DESCRIPTION
	[~pipelines.ledits_pp.LEditsPPInversionPipelineOutput]: Output will contain the resized input image(s)
	and respective VAE reconstruction(s).

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion.py

def invert(
    self,
    image: PipelineImageInput,
    source_prompt: str = "",
    source_guidance_scale: float = 3.5,
    num_inversion_steps: int = 30,
    skip: float = 0.15,
    generator: Optional[np.random.Generator] = None,
    cross_attention_kwargs: Optional[Dict[str, Any]] = None,
    clip_skip: Optional[int] = None,
    height: Optional[int] = None,
    width: Optional[int] = None,
    resize_mode: Optional[str] = "default",
    crops_coords: Optional[Tuple[int, int, int, int]] = None,
):
    r"""
    The function to the pipeline for image inversion as described by the [LEDITS++
    Paper](https://arxiv.org/abs/2301.12247). If the scheduler is set to [`~schedulers.DDIMScheduler`] the
    inversion proposed by [edit-friendly DPDM](https://arxiv.org/abs/2304.06140) will be performed instead.

    Args:
        image (`PipelineImageInput`):
            Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect
            ratio.
        source_prompt (`str`, defaults to `""`):
            Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled
            if the `source_prompt` is `""`.
        source_guidance_scale (`float`, defaults to `3.5`):
            Strength of guidance during inversion.
        num_inversion_steps (`int`, defaults to `30`):
            Number of total performed inversion steps after discarding the initial `skip` steps.
        skip (`float`, defaults to `0.15`):
            Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values
            will lead to stronger changes to the input image. `skip` has to be between `0` and `1`.
        generator (`np.random.Generator`, *optional*):
            A [`np.random.Generator`](https://numpy.org/doc/stable/reference/random/generator.html) to make inversion
            deterministic.
        cross_attention_kwargs (`dict`, *optional*):
            A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
            [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
        clip_skip (`int`, *optional*):
            Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
            the output of the pre-final layer will be used for computing the prompt embeddings.
        height (`int`, *optional*, defaults to `None`):
            The height in preprocessed image. If `None`, will use the `get_default_height_width()` to get default
            height.
        width (`int`, *optional*`, defaults to `None`):
            The width in preprocessed. If `None`, will use get_default_height_width()` to get the default width.
        resize_mode (`str`, *optional*, defaults to `default`):
            The resize mode, can be one of `default` or `fill`. If `default`, will resize the image to fit within
            the specified width and height, and it may not maintaining the original aspect ratio. If `fill`, will
            resize the image to fit within the specified width and height, maintaining the aspect ratio, and then
            center the image within the dimensions, filling empty with data from image. If `crop`, will resize the
            image to fit within the specified width and height, maintaining the aspect ratio, and then center the
            image within the dimensions, cropping the excess. Note that resize_mode `fill` and `crop` are only
            supported for PIL image input.
        crops_coords (`List[Tuple[int, int, int, int]]`, *optional*, defaults to `None`):
            The crop coordinates for each image in the batch. If `None`, will not crop the image.

    Returns:
        [`~pipelines.ledits_pp.LEditsPPInversionPipelineOutput`]: Output will contain the resized input image(s)
        and respective VAE reconstruction(s).
    """
    if height is not None and height % 32 != 0 or width is not None and width % 32 != 0:
        raise ValueError("height and width must be a factor of 32.")
    # Reset attn processor, we do not want to store attn maps during inversion
    self.unet.set_attn_processor(AttnProcessor())

    self.eta = 1.0

    self.scheduler.config.timestep_spacing = "leading"
    self.scheduler.set_timesteps(int(num_inversion_steps * (1 + skip)))
    self.inversion_steps = self.scheduler.timesteps[-num_inversion_steps:]
    timesteps = self.inversion_steps

    # 1. encode image
    x0, resized = self.encode_image(
        image,
        dtype=self.text_encoder.dtype,
        height=height,
        width=width,
        resize_mode=resize_mode,
        crops_coords=crops_coords,
    )
    self.batch_size = x0.shape[0]

    # autoencoder reconstruction
    image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]
    image_rec = self.image_processor.postprocess(image_rec, output_type="pil")

    # 2. get embeddings
    do_classifier_free_guidance = source_guidance_scale > 1.0

    lora_scale = cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None

    uncond_embedding, text_embeddings, _ = self.encode_prompt(
        num_images_per_prompt=1,
        negative_prompt=None,
        enable_edit_guidance=do_classifier_free_guidance,
        editing_prompt=source_prompt,
        lora_scale=lora_scale,
        clip_skip=clip_skip,
    )

    # 3. find zs and xts
    variance_noise_shape = (num_inversion_steps, *x0.shape)

    # intermediate latents
    t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
    xts = mint.zeros(size=variance_noise_shape, dtype=uncond_embedding.dtype)

    for t in reversed(timesteps):
        idx = num_inversion_steps - t_to_idx[int(t)] - 1
        noise = randn_tensor(shape=x0.shape, generator=generator, dtype=x0.dtype)
        xts[idx] = self.scheduler.add_noise(x0, noise, ms.tensor([t]))
    xts = mint.cat([x0.unsqueeze(0), xts], dim=0)

    self.scheduler.set_timesteps(len(self.scheduler.timesteps))
    # noise maps
    zs = mint.zeros(size=variance_noise_shape, dtype=uncond_embedding.dtype)

    with self.progress_bar(total=len(timesteps)) as progress_bar:
        for t in timesteps:
            idx = num_inversion_steps - t_to_idx[int(t)] - 1
            # 1. predict noise residual
            xt = xts[idx + 1]

            noise_pred = self.unet(xt, timestep=t, encoder_hidden_states=uncond_embedding)[0]

            if not source_prompt == "":
                noise_pred_cond = self.unet(xt, timestep=t, encoder_hidden_states=text_embeddings)[0]
                noise_pred = noise_pred + source_guidance_scale * (noise_pred_cond - noise_pred)

            xtm1 = xts[idx]
            z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, self.eta)
            zs[idx] = z

            # correction to avoid error accumulation
            xts[idx] = xtm1_corrected

            progress_bar.update()

    self.init_latents = xts[-1].broadcast_to((self.batch_size, -1, -1, -1))
    zs = zs.flip((0,))
    self.zs = zs

    return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec)

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL

Bases: DiffusionPipeline, FromSingleFileMixin, StableDiffusionXLLoraLoaderMixin, TextualInversionLoaderMixin, IPAdapterMixin

Pipeline for textual image editing using LEDits++ with Stable Diffusion XL.

This model inherits from [DiffusionPipeline] and builds on the [StableDiffusionXLPipeline]. Check the superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular device, etc.).

In addition the pipeline inherits the following loading methods

• LoRA: [LEditsPPPipelineStableDiffusionXL.load_lora_weights]
• Ckpt: [loaders.FromSingleFileMixin.from_single_file]

as well as the following saving methods

• LoRA: [loaders.StableDiffusionXLPipeline.save_lora_weights]

PARAMETER	DESCRIPTION
vae	Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. TYPE: [`AutoencoderKL`]
text_encoder	Frozen text-encoder. Stable Diffusion XL uses the text portion of CLIP, specifically the clip-vit-large-patch14 variant. TYPE: [`~transformers.CLIPTextModel`]
text_encoder_2	Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of CLIP, specifically the laion/CLIP-ViT-bigG-14-laion2B-39B-b160k variant. TYPE: [`~transformers.CLIPTextModelWithProjection`]
tokenizer	Tokenizer of class CLIPTokenizer. TYPE: [`~transformers.CLIPTokenizer`]
tokenizer_2	Second Tokenizer of class CLIPTokenizer. TYPE: [`~transformers.CLIPTokenizer`]
unet	Conditional U-Net architecture to denoise the encoded image latents. TYPE: [`UNet2DConditionModel`]
scheduler	A scheduler to be used in combination with unet to denoise the encoded image latens. Can be one of [DPMSolverMultistepScheduler] or [DDIMScheduler]. If any other scheduler is passed it will automatically be set to [DPMSolverMultistepScheduler]. TYPE: [`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]
force_zeros_for_empty_prompt	Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of stabilityai/stable-diffusion-xl-base-1-0. TYPE: `bool`, *optional*, defaults to `"True"` DEFAULT: True
add_watermarker	Whether to use the invisible_watermark library to watermark output images. If not defined, it will default to True if the package is installed, otherwise no watermarker will be used. TYPE: `bool`, *optional* DEFAULT: None

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

class LEditsPPPipelineStableDiffusionXL(
    DiffusionPipeline,
    FromSingleFileMixin,
    StableDiffusionXLLoraLoaderMixin,
    TextualInversionLoaderMixin,
    IPAdapterMixin,
):
    """
    Pipeline for textual image editing using LEDits++ with Stable Diffusion XL.

    This model inherits from [`DiffusionPipeline`] and builds on the [`StableDiffusionXLPipeline`]. Check the
    superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a
    particular device, etc.).

    In addition the pipeline inherits the following loading methods:
        - *LoRA*: [`LEditsPPPipelineStableDiffusionXL.load_lora_weights`]
        - *Ckpt*: [`loaders.FromSingleFileMixin.from_single_file`]

    as well as the following saving methods:
        - *LoRA*: [`loaders.StableDiffusionXLPipeline.save_lora_weights`]

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`~transformers.CLIPTextModel`]):
            Frozen text-encoder. Stable Diffusion XL uses the text portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        text_encoder_2 ([`~transformers.CLIPTextModelWithProjection`]):
            Second frozen text-encoder. Stable Diffusion XL uses the text and pool portion of
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModelWithProjection),
            specifically the
            [laion/CLIP-ViT-bigG-14-laion2B-39B-b160k](https://huggingface.co/laion/CLIP-ViT-bigG-14-laion2B-39B-b160k)
            variant.
        tokenizer ([`~transformers.CLIPTokenizer`]):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer_2 ([`~transformers.CLIPTokenizer`]):
            Second Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
        scheduler ([`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]):
            A scheduler to be used in combination with `unet` to denoise the encoded image latens. Can be one of
            [`DPMSolverMultistepScheduler`] or [`DDIMScheduler`]. If any other scheduler is passed it will
            automatically be set to [`DPMSolverMultistepScheduler`].
        force_zeros_for_empty_prompt (`bool`, *optional*, defaults to `"True"`):
            Whether the negative prompt embeddings shall be forced to always be set to 0. Also see the config of
            `stabilityai/stable-diffusion-xl-base-1-0`.
        add_watermarker (`bool`, *optional*):
            Whether to use the [invisible_watermark library](https://github.com/ShieldMnt/invisible-watermark/) to
            watermark output images. If not defined, it will default to True if the package is installed, otherwise no
            watermarker will be used.
    """

    model_cpu_offload_seq = "text_encoder->text_encoder_2->unet->vae"
    _optional_components = [
        "tokenizer",
        "tokenizer_2",
        "text_encoder",
        "text_encoder_2",
        "image_encoder",
        "feature_extractor",
    ]
    _callback_tensor_inputs = [
        "latents",
        "prompt_embeds",
        "negative_prompt_embeds",
        "add_text_embeds",
        "add_time_ids",
        "negative_pooled_prompt_embeds",
        "negative_add_time_ids",
    ]

    def __init__(
        self,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        text_encoder_2: CLIPTextModelWithProjection,
        tokenizer: CLIPTokenizer,
        tokenizer_2: CLIPTokenizer,
        unet: UNet2DConditionModel,
        scheduler: Union[DPMSolverMultistepScheduler, DDIMScheduler],
        image_encoder: CLIPVisionModelWithProjection = None,
        feature_extractor: CLIPImageProcessor = None,
        force_zeros_for_empty_prompt: bool = True,
        add_watermarker: Optional[bool] = None,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            unet=unet,
            scheduler=scheduler,
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
        )
        self.register_to_config(force_zeros_for_empty_prompt=force_zeros_for_empty_prompt)
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)

        if not isinstance(scheduler, DDIMScheduler) and not isinstance(scheduler, DPMSolverMultistepScheduler):
            self.scheduler = DPMSolverMultistepScheduler.from_config(
                scheduler.config, algorithm_type="sde-dpmsolver++", solver_order=2
            )
            logger.warning(
                "This pipeline only supports DDIMScheduler and DPMSolverMultistepScheduler. "
                "The scheduler has been changed to DPMSolverMultistepScheduler."
            )

        self.default_sample_size = (
            self.unet.config.sample_size
            if hasattr(self, "unet") and self.unet is not None and hasattr(self.unet.config, "sample_size")
            else 128
        )

        add_watermarker = add_watermarker if add_watermarker is not None else is_invisible_watermark_available()

        if add_watermarker:
            self.watermark = StableDiffusionXLWatermarker()
        else:
            self.watermark = None
        self.inversion_steps = None

    def encode_prompt(
        self,
        num_images_per_prompt: int = 1,
        negative_prompt: Optional[str] = None,
        negative_prompt_2: Optional[str] = None,
        negative_prompt_embeds: Optional[ms.Tensor] = None,
        negative_pooled_prompt_embeds: Optional[ms.Tensor] = None,
        lora_scale: Optional[float] = None,
        clip_skip: Optional[int] = None,
        enable_edit_guidance: bool = True,
        editing_prompt: Optional[str] = None,
        editing_prompt_embeds: Optional[ms.Tensor] = None,
        editing_pooled_prompt_embeds: Optional[ms.Tensor] = None,
    ) -> object:
        r"""
        Encodes the prompt into text encoder hidden states.

        Args:
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead.
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            negative_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            negative_pooled_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            enable_edit_guidance (`bool`):
                Whether to guide towards an editing prompt or not.
            editing_prompt (`str` or `List[str]`, *optional*):
                Editing prompt(s) to be encoded. If not defined and 'enable_edit_guidance' is True, one has to pass
                `editing_prompt_embeds` instead.
            editing_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided and 'enable_edit_guidance' is True, editing_prompt_embeds will be generated from
                `editing_prompt` input argument.
            editing_pooled_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated edit pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled editing_pooled_prompt_embeds will be generated from `editing_prompt`
                input argument.
        """
        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if self.text_encoder is not None:
                scale_lora_layers(self.text_encoder, lora_scale)

            if self.text_encoder_2 is not None:
                scale_lora_layers(self.text_encoder_2, lora_scale)

        batch_size = self.batch_size

        # Define tokenizers and text encoders
        tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
        text_encoders = (
            [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
        )
        num_edit_tokens = 0

        # get unconditional embeddings for classifier free guidance
        zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt

        if negative_prompt_embeds is None:
            negative_prompt = negative_prompt or ""
            negative_prompt_2 = negative_prompt_2 or negative_prompt

            # normalize str to list
            negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
            negative_prompt_2 = (
                batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
            )

            uncond_tokens: List[str]

            if batch_size != len(negative_prompt):
                raise ValueError(
                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but image inversion "
                    f" has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                    " the batch size of the input images."
                )
            else:
                uncond_tokens = [negative_prompt, negative_prompt_2]

            negative_prompt_embeds_list = []
            for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)

                uncond_input = tokenizer(
                    negative_prompt,
                    padding="max_length",
                    max_length=tokenizer.model_max_length,
                    truncation=True,
                    return_tensors="np",
                )

                negative_prompt_embeds = text_encoder(
                    ms.tensor(uncond_input.input_ids),
                    output_hidden_states=True,
                )
                # We are only ALWAYS interested in the pooled output of the final text encoder
                negative_pooled_prompt_embeds = negative_prompt_embeds[0]
                negative_prompt_embeds = negative_prompt_embeds[2][-2]

                negative_prompt_embeds_list.append(negative_prompt_embeds)

            negative_prompt_embeds = mint.concat(negative_prompt_embeds_list, dim=-1)

            if zero_out_negative_prompt:
                negative_prompt_embeds = mint.zeros_like(negative_prompt_embeds)
                negative_pooled_prompt_embeds = mint.zeros_like(negative_pooled_prompt_embeds)

        if enable_edit_guidance and editing_prompt_embeds is None:
            editing_prompt_2 = editing_prompt

            editing_prompts = [editing_prompt, editing_prompt_2]
            edit_prompt_embeds_list = []

            for editing_prompt, tokenizer, text_encoder in zip(editing_prompts, tokenizers, text_encoders):
                if isinstance(self, TextualInversionLoaderMixin):
                    editing_prompt = self.maybe_convert_prompt(editing_prompt, tokenizer)

                max_length = negative_prompt_embeds.shape[1]
                edit_concepts_input = tokenizer(
                    # [x for item in editing_prompt for x in repeat(item, batch_size)],
                    editing_prompt,
                    padding="max_length",
                    max_length=max_length,
                    truncation=True,
                    return_tensors="np",
                    return_length=True,
                )
                num_edit_tokens = edit_concepts_input.length - 2

                edit_concepts_embeds = text_encoder(
                    ms.tensor(edit_concepts_input.input_ids),
                    output_hidden_states=True,
                )
                # We are only ALWAYS interested in the pooled output of the final text encoder
                editing_pooled_prompt_embeds = edit_concepts_embeds[0]
                if clip_skip is None:
                    edit_concepts_embeds = edit_concepts_embeds[2][-2]
                else:
                    # "2" because SDXL always indexes from the penultimate layer.
                    edit_concepts_embeds = edit_concepts_embeds[-2][-(clip_skip + 2)]

                edit_prompt_embeds_list.append(edit_concepts_embeds)

            edit_concepts_embeds = mint.concat(edit_prompt_embeds_list, dim=-1)
        elif not enable_edit_guidance:
            edit_concepts_embeds = None
            editing_pooled_prompt_embeds = None

        negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype)
        bs_embed, seq_len, _ = negative_prompt_embeds.shape
        # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
        seq_len = negative_prompt_embeds.shape[1]
        negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype)
        negative_prompt_embeds = negative_prompt_embeds.tile((1, num_images_per_prompt, 1))
        negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        if enable_edit_guidance:
            bs_embed_edit, seq_len, _ = edit_concepts_embeds.shape
            edit_concepts_embeds = edit_concepts_embeds.to(dtype=self.text_encoder_2.dtype)
            edit_concepts_embeds = edit_concepts_embeds.tile((1, num_images_per_prompt, 1))
            edit_concepts_embeds = edit_concepts_embeds.view(bs_embed_edit * num_images_per_prompt, seq_len, -1)

        negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.tile((1, num_images_per_prompt)).view(
            bs_embed * num_images_per_prompt, -1
        )

        if enable_edit_guidance:
            editing_pooled_prompt_embeds = editing_pooled_prompt_embeds.tile((1, num_images_per_prompt)).view(
                bs_embed_edit * num_images_per_prompt, -1
            )

        if self.text_encoder is not None:
            if isinstance(self, StableDiffusionXLLoraLoaderMixin):
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder, lora_scale)

        if self.text_encoder_2 is not None:
            if isinstance(self, StableDiffusionXLLoraLoaderMixin):
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)

        return (
            negative_prompt_embeds,
            edit_concepts_embeds,
            negative_pooled_prompt_embeds,
            editing_pooled_prompt_embeds,
            num_edit_tokens,
        )

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
    def prepare_extra_step_kwargs(self, eta, generator=None):
        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
        # and should be between [0, 1]

        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
        extra_step_kwargs = {}
        if accepts_eta:
            extra_step_kwargs["eta"] = eta

        # check if the scheduler accepts generator
        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
        if accepts_generator:
            extra_step_kwargs["generator"] = generator
        return extra_step_kwargs

    def check_inputs(
        self,
        negative_prompt=None,
        negative_prompt_2=None,
        negative_prompt_embeds=None,
        negative_pooled_prompt_embeds=None,
    ):
        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )
        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
            raise ValueError(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."  # noqa: E501
            )

    # Modified from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
    def prepare_latents(self, latents):
        dtype = latents.dtype

        # scale the initial noise by the standard deviation required by the scheduler
        latents = (latents * self.scheduler.init_noise_sigma).to(dtype)
        return latents

    def _get_add_time_ids(
        self, original_size, crops_coords_top_left, target_size, dtype, text_encoder_projection_dim=None
    ):
        add_time_ids = list(original_size + crops_coords_top_left + target_size)

        passed_add_embed_dim = (
            self.unet.config.addition_time_embed_dim * len(add_time_ids) + text_encoder_projection_dim
        )
        expected_add_embed_dim = self.unet.add_embedding.linear_1.in_features

        if expected_add_embed_dim != passed_add_embed_dim:
            raise ValueError(
                f"Model expects an added time embedding vector of length {expected_add_embed_dim}, but a vector of {passed_add_embed_dim} was created. The model has an incorrect config. Please check `unet.config.time_embedding_type` and `text_encoder_2.config.projection_dim`."  # noqa: E501
            )

        add_time_ids = ms.tensor([add_time_ids], dtype=dtype)
        return add_time_ids

    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_upscale.StableDiffusionUpscalePipeline.upcast_vae
    def upcast_vae(self):
        self.vae.to(dtype=ms.float32)

    # Copied from diffusers.pipelines.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
    def get_guidance_scale_embedding(
        self, w: ms.Tensor, embedding_dim: int = 512, dtype: ms.Type = ms.float32
    ) -> ms.Tensor:
        """
        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

        Args:
            w (`ms.Tensor`):
                Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
            embedding_dim (`int`, *optional*, defaults to 512):
                Dimension of the embeddings to generate.
            dtype (`ms.Type`, *optional*, defaults to `ms.float32`):
                Data type of the generated embeddings.

        Returns:
            `ms.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
        """
        assert len(w.shape) == 1
        w = w * 1000.0

        half_dim = embedding_dim // 2
        emb = mint.log(ms.tensor(10000.0)) / (half_dim - 1)
        emb = mint.exp(mint.arange(half_dim, dtype=dtype) * -emb)
        emb = w.to(dtype)[:, None] * emb[None, :]
        emb = mint.cat([mint.sin(emb), mint.cos(emb)], dim=1)
        if embedding_dim % 2 == 1:  # zero pad
            emb = mint.nn.functional.pad(emb, (0, 1))
        assert emb.shape == (w.shape[0], embedding_dim)
        return emb

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def guidance_rescale(self):
        return self._guidance_rescale

    @property
    def clip_skip(self):
        return self._clip_skip

    # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
    # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
    # corresponds to doing no classifier free guidance.
    @property
    def do_classifier_free_guidance(self):
        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None

    @property
    def cross_attention_kwargs(self):
        return self._cross_attention_kwargs

    @property
    def denoising_end(self):
        return self._denoising_end

    @property
    def num_timesteps(self):
        return self._num_timesteps

    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    # Copied from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEditsPPPipelineStableDiffusion.prepare_unet
    def prepare_unet(self, attention_store, PnP: bool = False):
        attn_procs = {}
        for name in self.unet.attn_processors.keys():
            if name.startswith("mid_block"):
                place_in_unet = "mid"
            elif name.startswith("up_blocks"):
                place_in_unet = "up"
            elif name.startswith("down_blocks"):
                place_in_unet = "down"
            else:
                continue

            if "attn2" in name and place_in_unet != "mid":
                attn_procs[name] = LEDITSCrossAttnProcessor(
                    attention_store=attention_store,
                    place_in_unet=place_in_unet,
                    pnp=PnP,
                    editing_prompts=self.enabled_editing_prompts,
                )
            else:
                attn_procs[name] = AttnProcessor()

        self.unet.set_attn_processor(attn_procs)

    def __call__(
        self,
        denoising_end: Optional[float] = None,
        negative_prompt: Optional[Union[str, List[str]]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        negative_prompt_embeds: Optional[ms.Tensor] = None,
        negative_pooled_prompt_embeds: Optional[ms.Tensor] = None,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = False,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        guidance_rescale: float = 0.0,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        target_size: Optional[Tuple[int, int]] = None,
        editing_prompt: Optional[Union[str, List[str]]] = None,
        editing_prompt_embeddings: Optional[ms.Tensor] = None,
        editing_pooled_prompt_embeds: Optional[ms.Tensor] = None,
        reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
        edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
        edit_warmup_steps: Optional[Union[int, List[int]]] = 0,
        edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
        edit_threshold: Optional[Union[float, List[float]]] = 0.9,
        sem_guidance: Optional[List[ms.Tensor]] = None,
        use_cross_attn_mask: bool = False,
        use_intersect_mask: bool = False,
        user_mask: Optional[ms.Tensor] = None,
        attn_store_steps: Optional[List[int]] = [],
        store_averaged_over_steps: bool = True,
        clip_skip: Optional[int] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        **kwargs,
    ):
        r"""
        The call function to the pipeline for editing. The
        [`~pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.invert`] method has to be called beforehand. Edits
        will always be performed for the last inverted image(s).

        Args:
            denoising_end (`float`, *optional*):
                When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
                completed before it is intentionally prematurely terminated. As a result, the returned sample will
                still retain a substantial amount of noise as determined by the discrete timesteps selected by the
                scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
                "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            negative_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            negative_pooled_prompt_embeds (`ms.Tensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
                input argument.
            ip_adapter_image: (`PipelineImageInput`, *optional*):
                Optional image input to work with IP Adapters.
            output_type (`str`, *optional*, defaults to `"pil"`):
                The output format of the generate image. Choose between
                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
            return_dict (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
                of a plain tuple.
            callback (`Callable`, *optional*):
                A function that will be called every `callback_steps` steps during inference. The function will be
                called with the following arguments: `callback(step: int, timestep: int, latents: ms.Tensor)`.
            callback_steps (`int`, *optional*, defaults to 1):
                The frequency at which the `callback` function will be called. If not specified, the callback will be
                called at every step.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
            guidance_rescale (`float`, *optional*, defaults to 0.7):
                Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
                Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
                [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
                Guidance rescale factor should fix overexposure when using zero terminal SNR.
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
                For most cases, `target_size` should be set to the desired height and width of the generated image. If
                not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
                section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            editing_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. The image is reconstructed by setting
                `editing_prompt = None`. Guidance direction of prompt should be specified via
                `reverse_editing_direction`.
            editing_prompt_embeddings (`ms.Tensor`, *optional*):
                Pre-generated edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
                If not provided, editing_prompt_embeddings will be generated from `editing_prompt` input argument.
            editing_pooled_prompt_embeddings (`ms.Tensor`, *optional*):
                Pre-generated pooled edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, editing_prompt_embeddings will be generated from `editing_prompt` input
                argument.
            reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
                Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
            edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
                Guidance scale for guiding the image generation. If provided as list values should correspond to
                `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 12 of [LEDITS++
                Paper](https://arxiv.org/abs/2301.12247).
            edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
                Number of diffusion steps (for each prompt) for which guidance is not applied.
            edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
                Number of diffusion steps (for each prompt) after which guidance is no longer applied.
            edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
                Masking threshold of guidance. Threshold should be proportional to the image region that is modified.
                'edit_threshold' is defined as 'λ' of equation 12 of [LEDITS++
                Paper](https://arxiv.org/abs/2301.12247).
            sem_guidance (`List[ms.Tensor]`, *optional*):
                List of pre-generated guidance vectors to be applied at generation. Length of the list has to
                correspond to `num_inference_steps`.
            use_cross_attn_mask:
                Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask
                is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of [LEDITS++
                paper](https://arxiv.org/pdf/2311.16711.pdf).
            use_intersect_mask:
                Whether the masking term is calculated as intersection of cross-attention masks and masks derived from
                the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate
                are defined as 'M^2' of equation 12 of [LEDITS++ paper](https://arxiv.org/pdf/2311.16711.pdf).
            user_mask:
                User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s
                implicit masks do not meet user preferences.
            attn_store_steps:
                Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes.
            store_averaged_over_steps:
                Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If
                False, attention maps for each step are stores separately. Just for visualization purposes.
            clip_skip (`int`, *optional*):
                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
                the output of the pre-final layer will be used for computing the prompt embeddings.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference. The function is called
                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
                `callback_on_step_end_tensor_inputs`.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
                `._callback_tensor_inputs` attribute of your pipeline class.

        Examples:

        Returns:
            [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] or `tuple`:
            [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
            returning a tuple, the first element is a list with the generated images.
        """
        if self.inversion_steps is None:
            raise ValueError(
                "You need to invert an input image first before calling the pipeline. The `invert` method has to be called beforehand. Edits will always be performed for the last inverted image(s)."  # noqa: E501
            )

        eta = self.eta
        num_images_per_prompt = 1
        latents = self.init_latents

        zs = self.zs
        self.scheduler.set_timesteps(len(self.scheduler.timesteps))

        if use_intersect_mask:
            use_cross_attn_mask = True

        if use_cross_attn_mask:
            self.smoothing = LeditsGaussianSmoothing()

        # TODO: Check inputs
        # 1. Check inputs. Raise error if not correct
        # self.check_inputs(
        #    callback_steps,
        #    negative_prompt,
        #    negative_prompt_2,
        #    prompt_embeds,
        #    negative_prompt_embeds,
        #    pooled_prompt_embeds,
        #    negative_pooled_prompt_embeds,
        # )
        self._guidance_rescale = guidance_rescale
        self._clip_skip = clip_skip
        self._cross_attention_kwargs = cross_attention_kwargs
        self._denoising_end = denoising_end

        # 2. Define call parameters
        batch_size = self.batch_size

        if editing_prompt:
            enable_edit_guidance = True
            if isinstance(editing_prompt, str):
                editing_prompt = [editing_prompt]
            self.enabled_editing_prompts = len(editing_prompt)
        elif editing_prompt_embeddings is not None:
            enable_edit_guidance = True
            self.enabled_editing_prompts = editing_prompt_embeddings.shape[0]
        else:
            self.enabled_editing_prompts = 0
            enable_edit_guidance = False

        # 3. Encode input prompt
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )
        (
            prompt_embeds,
            edit_prompt_embeds,
            negative_pooled_prompt_embeds,
            pooled_edit_embeds,
            num_edit_tokens,
        ) = self.encode_prompt(
            num_images_per_prompt=num_images_per_prompt,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            negative_prompt_embeds=negative_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            lora_scale=text_encoder_lora_scale,
            clip_skip=self.clip_skip,
            enable_edit_guidance=enable_edit_guidance,
            editing_prompt=editing_prompt,
            editing_prompt_embeds=editing_prompt_embeddings,
            editing_pooled_prompt_embeds=editing_pooled_prompt_embeds,
        )

        # 4. Prepare timesteps
        # self.scheduler.set_timesteps(num_inference_steps)

        timesteps = self.inversion_steps
        t_to_idx = {int(v): k for k, v in enumerate(timesteps)}

        if use_cross_attn_mask:
            self.attention_store = LeditsAttentionStore(
                average=store_averaged_over_steps,
                batch_size=batch_size,
                max_size=(latents.shape[-2] / 4.0) * (latents.shape[-1] / 4.0),
                max_resolution=None,
            )
            self.prepare_unet(self.attention_store)
            resolution = latents.shape[-2:]
            att_res = (int(resolution[0] / 4), int(resolution[1] / 4))

        # 5. Prepare latent variables
        latents = self.prepare_latents(latents=latents)

        # 6. Prepare extra step kwargs.
        extra_step_kwargs = self.prepare_extra_step_kwargs(eta)

        if self.text_encoder_2 is None:
            text_encoder_projection_dim = int(negative_pooled_prompt_embeds.shape[-1])
        else:
            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

        # 7. Prepare added time ids & embeddings
        add_text_embeds = negative_pooled_prompt_embeds
        add_time_ids = self._get_add_time_ids(
            self.size,
            crops_coords_top_left,
            self.size,
            dtype=negative_pooled_prompt_embeds.dtype,
            text_encoder_projection_dim=text_encoder_projection_dim,
        )

        if enable_edit_guidance:
            prompt_embeds = mint.cat([prompt_embeds, edit_prompt_embeds], dim=0)
            add_text_embeds = mint.cat([add_text_embeds, pooled_edit_embeds], dim=0)
            edit_concepts_time_ids = add_time_ids.tile((edit_prompt_embeds.shape[0], 1))
            add_time_ids = mint.cat([add_time_ids, edit_concepts_time_ids], dim=0)
            self.text_cross_attention_maps = [editing_prompt] if isinstance(editing_prompt, str) else editing_prompt

        prompt_embeds = prompt_embeds
        add_text_embeds = add_text_embeds
        add_time_ids = add_time_ids.tile((batch_size * num_images_per_prompt, 1))

        if ip_adapter_image is not None:
            # TODO: fix image encoding
            image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, num_images_per_prompt)
            if self.do_classifier_free_guidance:
                image_embeds = mint.cat([negative_image_embeds, image_embeds])
                image_embeds = image_embeds

        # 8. Denoising loop
        self.sem_guidance = None
        self.activation_mask = None

        if (
            self.denoising_end is not None
            and isinstance(self.denoising_end, float)
            and self.denoising_end > 0
            and self.denoising_end < 1
        ):
            discrete_timestep_cutoff = int(
                round(
                    self.scheduler.config.num_train_timesteps
                    - (self.denoising_end * self.scheduler.config.num_train_timesteps)
                )
            )
            num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
            timesteps = timesteps[:num_inference_steps]

        # 9. Optionally get Guidance Scale Embedding
        timestep_cond = None
        if self.unet.config.time_cond_proj_dim is not None:
            guidance_scale_tensor = ms.tensor(self.guidance_scale - 1).tile((batch_size * num_images_per_prompt))
            timestep_cond = self.get_guidance_scale_embedding(
                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
            ).to(dtype=latents.dtype)

        self._num_timesteps = len(timesteps)
        with self.progress_bar(total=self._num_timesteps) as progress_bar:
            for i, t in enumerate(timesteps):
                # expand the latents if we are doing classifier free guidance
                latent_model_input = mint.cat([latents] * (1 + self.enabled_editing_prompts))
                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
                # predict the noise residual
                added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
                if ip_adapter_image is not None:
                    added_cond_kwargs["image_embeds"] = image_embeds
                noise_pred = self.unet(
                    latent_model_input,
                    t,
                    encoder_hidden_states=prompt_embeds,
                    cross_attention_kwargs=cross_attention_kwargs,
                    added_cond_kwargs=added_cond_kwargs,
                    return_dict=False,
                )[0]

                noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts)  # [b,4, 64, 64]
                noise_pred_uncond = noise_pred_out[0]
                noise_pred_edit_concepts = noise_pred_out[1:]

                noise_guidance_edit = mint.zeros(
                    noise_pred_uncond.shape,
                    dtype=noise_pred_uncond.dtype,
                )

                if sem_guidance is not None and len(sem_guidance) > i:
                    noise_guidance_edit += sem_guidance[i]

                elif enable_edit_guidance:
                    if self.activation_mask is None:
                        self.activation_mask = mint.zeros(
                            (len(timesteps), self.enabled_editing_prompts, *noise_pred_edit_concepts[0].shape)
                        )
                    if self.sem_guidance is None:
                        self.sem_guidance = mint.zeros((len(timesteps), *noise_pred_uncond.shape))

                    # noise_guidance_edit = mint.zeros_like(noise_guidance)
                    for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
                        if isinstance(edit_warmup_steps, list):
                            edit_warmup_steps_c = edit_warmup_steps[c]
                        else:
                            edit_warmup_steps_c = edit_warmup_steps
                        if i < edit_warmup_steps_c:
                            continue

                        if isinstance(edit_guidance_scale, list):
                            edit_guidance_scale_c = edit_guidance_scale[c]
                        else:
                            edit_guidance_scale_c = edit_guidance_scale

                        if isinstance(edit_threshold, list):
                            edit_threshold_c = edit_threshold[c]
                        else:
                            edit_threshold_c = edit_threshold
                        if isinstance(reverse_editing_direction, list):
                            reverse_editing_direction_c = reverse_editing_direction[c]
                        else:
                            reverse_editing_direction_c = reverse_editing_direction

                        if isinstance(edit_cooldown_steps, list):
                            edit_cooldown_steps_c = edit_cooldown_steps[c]
                        elif edit_cooldown_steps is None:
                            edit_cooldown_steps_c = i + 1
                        else:
                            edit_cooldown_steps_c = edit_cooldown_steps

                        if i >= edit_cooldown_steps_c:
                            continue

                        noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond

                        if reverse_editing_direction_c:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1

                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c

                        if user_mask is not None:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask

                        if use_cross_attn_mask:
                            out = self.attention_store.aggregate_attention(
                                attention_maps=self.attention_store.step_store,
                                prompts=self.text_cross_attention_maps,
                                res=att_res,
                                from_where=["up", "down"],
                                is_cross=True,
                                select=self.text_cross_attention_maps.index(editing_prompt[c]),
                            )
                            attn_map = out[:, :, :, 1 : 1 + num_edit_tokens[c]]  # 0 -> startoftext

                            # average over all tokens
                            if attn_map.shape[3] != num_edit_tokens[c]:
                                raise ValueError(
                                    f"Incorrect shape of attention_map. Expected size {num_edit_tokens[c]}, but found {attn_map.shape[3]}!"
                                )
                            attn_map = mint.sum(attn_map, dim=3)

                            # gaussian_smoothing
                            attn_map = mint.nn.functional.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect")
                            attn_map = self.smoothing(attn_map).squeeze(1)

                            # ops.quantile function expects float32
                            # TODO: ops.quantile is not supported
                            if attn_map.dtype == ms.float32:
                                tmp = ms.tensor(
                                    np.quantile(attn_map.flatten(start_dim=1).numpy(), edit_threshold_c, axis=1)
                                )
                            else:
                                tmp = ms.tensor(
                                    np.quantile(
                                        attn_map.flatten(start_dim=1).to(ms.float32).numpy(), edit_threshold_c, axis=1
                                    )
                                ).to(attn_map.dtype)
                            attn_mask = mint.where(
                                attn_map >= tmp.unsqueeze(1).unsqueeze(1).tile((1, *att_res)), 1.0, 0.0
                            )

                            # resolution must match latent space dimension
                            attn_mask = mint.nn.functional.interpolate(
                                attn_mask.unsqueeze(1),
                                noise_guidance_edit_tmp.shape[-2:],  # 64,64
                            ).tile((1, 4, 1, 1))
                            self.activation_mask[i, c] = ops.stop_gradient(attn_mask)
                            if not use_intersect_mask:
                                noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                        if use_intersect_mask:
                            noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                            noise_guidance_edit_tmp_quantile = mint.sum(
                                noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                            )
                            noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile(
                                (1, self.unet.config.in_channels, 1, 1)
                            )

                            # ops.quantile function expects float32
                            # TODO: ops.quantile is not supported
                            if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                                tmp = ms.tensor(
                                    np.quantile(
                                        noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                        edit_threshold_c,
                                        axis=2,
                                        keepdims=False,
                                    )
                                )
                            else:
                                tmp = ms.tensor(
                                    np.quantile(
                                        noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(ms.float32).numpy(),
                                        edit_threshold_c,
                                        axis=2,
                                        keepdims=False,
                                    )
                                ).to(noise_guidance_edit_tmp_quantile.dtype)

                            intersect_mask = (
                                mint.where(
                                    noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                    mint.ones_like(noise_guidance_edit_tmp),
                                    mint.zeros_like(noise_guidance_edit_tmp),
                                )
                                * attn_mask
                            )

                            self.activation_mask[i, c] = ops.stop_gradient(intersect_mask)

                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask

                        elif not use_cross_attn_mask:
                            # calculate quantile
                            noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                            noise_guidance_edit_tmp_quantile = mint.sum(
                                noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                            )
                            noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile((1, 4, 1, 1))

                            # ops.quantile function expects float32
                            # TODO: ops.quantile is not supported
                            if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                                tmp = ms.tensor(
                                    np.quantile(
                                        noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                        edit_threshold_c,
                                        axis=2,
                                        keepdims=False,
                                    )
                                )
                            else:
                                tmp = ms.tensor(
                                    np.quantile(
                                        noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(ms.float32).numpy(),
                                        edit_threshold_c,
                                        axis=2,
                                        keepdims=False,
                                    )
                                ).to(noise_guidance_edit_tmp_quantile.dtype)

                            self.activation_mask[i, c] = ops.stop_gradient(
                                mint.where(
                                    noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                    mint.ones_like(noise_guidance_edit_tmp),
                                    mint.zeros_like(noise_guidance_edit_tmp),
                                )
                            )

                            noise_guidance_edit_tmp = mint.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                noise_guidance_edit_tmp,
                                mint.zeros_like(noise_guidance_edit_tmp),
                            )

                        noise_guidance_edit += noise_guidance_edit_tmp

                    self.sem_guidance[i] = ops.stop_gradient(noise_guidance_edit)

                noise_pred = noise_pred_uncond + noise_guidance_edit

                # compute the previous noisy sample x_t -> x_t-1
                if enable_edit_guidance and self.guidance_rescale > 0.0:
                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                    noise_pred = rescale_noise_cfg(
                        noise_pred,
                        noise_pred_edit_concepts.mean(dim=0, keepdim=False),
                        guidance_rescale=self.guidance_rescale,
                    )

                idx = t_to_idx[int(t)]
                latents = self.scheduler.step(
                    noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs, return_dict=False
                )[0]

                # step callback
                if use_cross_attn_mask:
                    store_step = i in attn_store_steps
                    self.attention_store.between_steps(store_step)

                if callback_on_step_end is not None:
                    callback_kwargs = {}
                    for k in callback_on_step_end_tensor_inputs:
                        callback_kwargs[k] = locals()[k]
                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                    latents = callback_outputs.pop("latents", latents)
                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
                    add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
                    negative_pooled_prompt_embeds = callback_outputs.pop(
                        "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
                    )
                    add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
                    # negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)

                # call the callback, if provided
                if i == len(timesteps) - 1 or ((i + 1) > 0 and (i + 1) % self.scheduler.order == 0):
                    progress_bar.update()

        if not output_type == "latent":
            # make sure the VAE is in float32 mode, as it overflows in float16
            needs_upcasting = self.vae.dtype == ms.float16 and self.vae.config.force_upcast

            if needs_upcasting:
                self.upcast_vae()
                latents = latents.to(next(iter(self.vae.post_quant_conv.get_parameters())).dtype)

            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]

            # cast back to fp16 if needed
            if needs_upcasting:
                self.vae.to(dtype=ms.float16)
        else:
            image = latents

        if not output_type == "latent":
            # apply watermark if available
            if self.watermark is not None:
                image = self.watermark.apply_watermark(image)

            image = self.image_processor.postprocess(image, output_type=output_type)

        if not return_dict:
            return (image,)

        return LEditsPPDiffusionPipelineOutput(images=image, nsfw_content_detected=None)

    # Modified from diffusers.pipelines.ledits_pp.pipeline_leditspp_stable_diffusion.LEditsPPPipelineStableDiffusion.encode_image
    def encode_image(self, image, dtype=None, height=None, width=None, resize_mode="default", crops_coords=None):
        image = self.image_processor.preprocess(
            image=image, height=height, width=width, resize_mode=resize_mode, crops_coords=crops_coords
        )
        height, width = image.shape[-2:]
        if height % 32 != 0 or width % 32 != 0:
            raise ValueError(
                "Image height and width must be a factor of 32. "
                "Consider down-sampling the input using the `height` and `width` parameters"
            )
        resized = self.image_processor.postprocess(image=image, output_type="pil")

        if max(image.shape[-2:]) > self.vae.config["sample_size"] * 1.5:
            logger.warning(
                "Your input images far exceed the default resolution of the underlying diffusion model. "
                "The output images may contain severe artifacts! "
                "Consider down-sampling the input using the `height` and `width` parameters"
            )
        image = image.to(dtype=dtype)
        needs_upcasting = self.vae.dtype == ms.float16 and self.vae.config.force_upcast

        if needs_upcasting:
            image = image.float()
            self.upcast_vae()

        x0 = self.vae.diag_gauss_dist.mode(self.vae.encode(image)[0])
        x0 = x0.to(dtype)
        # cast back to fp16 if needed
        if needs_upcasting:
            self.vae.to(dtype=ms.float16)

        x0 = self.vae.config.scaling_factor * x0
        return x0, resized

    def invert(
        self,
        image: PipelineImageInput,
        source_prompt: str = "",
        source_guidance_scale=3.5,
        negative_prompt: str = None,
        negative_prompt_2: str = None,
        num_inversion_steps: int = 50,
        skip: float = 0.15,
        generator: Optional[np.random.Generator] = None,
        crops_coords_top_left: Tuple[int, int] = (0, 0),
        num_zero_noise_steps: int = 3,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        resize_mode: Optional[str] = "default",
        crops_coords: Optional[Tuple[int, int, int, int]] = None,
    ):
        r"""
        The function to the pipeline for image inversion as described by the [LEDITS++
        Paper](https://arxiv.org/abs/2301.12247). If the scheduler is set to [`~schedulers.DDIMScheduler`] the
        inversion proposed by [edit-friendly DPDM](https://arxiv.org/abs/2304.06140) will be performed instead.

        Args:
            image (`PipelineImageInput`):
                Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect
                ratio.
            source_prompt (`str`, defaults to `""`):
                Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled
                if the `source_prompt` is `""`.
            source_guidance_scale (`float`, defaults to `3.5`):
                Strength of guidance during inversion.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
                less than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
            num_inversion_steps (`int`, defaults to `50`):
                Number of total performed inversion steps after discarding the initial `skip` steps.
            skip (`float`, defaults to `0.15`):
                Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values
                will lead to stronger changes to the input image. `skip` has to be between `0` and `1`.
            generator (`np.random.Generator`, *optional*):
                A [`np.random.Generator`](https://numpy.org/doc/stable/reference/random/generator.html) to make inversion
                deterministic.
            crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
                `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
                `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
                `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
                [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
            num_zero_noise_steps (`int`, defaults to `3`):
                Number of final diffusion steps that will not renoise the current image. If no steps are set to zero
                SD-XL in combination with [`DPMSolverMultistepScheduler`] will produce noise artifacts.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
                `self.processor` in
                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).

        Returns:
            [`~pipelines.ledits_pp.LEditsPPInversionPipelineOutput`]: Output will contain the resized input image(s)
            and respective VAE reconstruction(s).
        """
        if height is not None and height % 32 != 0 or width is not None and width % 32 != 0:
            raise ValueError("height and width must be a factor of 32.")

        # Reset attn processor, we do not want to store attn maps during inversion
        self.unet.set_attn_processor(AttnProcessor())

        self.eta = 1.0

        self.scheduler.config.timestep_spacing = "leading"
        self.scheduler.set_timesteps(int(num_inversion_steps * (1 + skip)))
        self.inversion_steps = self.scheduler.timesteps[-num_inversion_steps:]
        timesteps = self.inversion_steps

        num_images_per_prompt = 1

        # 0. Ensure that only uncond embedding is used if prompt = ""
        if source_prompt == "":
            # noise pred should only be noise_pred_uncond
            source_guidance_scale = 0.0
            do_classifier_free_guidance = False
        else:
            do_classifier_free_guidance = source_guidance_scale > 1.0

        # 1. prepare image
        x0, resized = self.encode_image(
            image,
            dtype=self.text_encoder_2.dtype,
            height=height,
            width=width,
            resize_mode=resize_mode,
            crops_coords=crops_coords,
        )
        width = x0.shape[2] * self.vae_scale_factor
        height = x0.shape[3] * self.vae_scale_factor
        self.size = (height, width)

        self.batch_size = x0.shape[0]

        # 2. get embeddings
        text_encoder_lora_scale = (
            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
        )

        if isinstance(source_prompt, str):
            source_prompt = [source_prompt] * self.batch_size

        (
            negative_prompt_embeds,
            prompt_embeds,
            negative_pooled_prompt_embeds,
            edit_pooled_prompt_embeds,
            _,
        ) = self.encode_prompt(
            num_images_per_prompt=num_images_per_prompt,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            editing_prompt=source_prompt,
            lora_scale=text_encoder_lora_scale,
            enable_edit_guidance=do_classifier_free_guidance,
        )
        if self.text_encoder_2 is None:
            text_encoder_projection_dim = int(negative_pooled_prompt_embeds.shape[-1])
        else:
            text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

        # 3. Prepare added time ids & embeddings
        add_text_embeds = negative_pooled_prompt_embeds
        add_time_ids = self._get_add_time_ids(
            self.size,
            crops_coords_top_left,
            self.size,
            dtype=negative_prompt_embeds.dtype,
            text_encoder_projection_dim=text_encoder_projection_dim,
        )

        if do_classifier_free_guidance:
            negative_prompt_embeds = mint.cat([negative_prompt_embeds, prompt_embeds], dim=0)
            add_text_embeds = mint.cat([add_text_embeds, edit_pooled_prompt_embeds], dim=0)
            add_time_ids = mint.cat([add_time_ids, add_time_ids], dim=0)

        negative_prompt_embeds = negative_prompt_embeds

        add_text_embeds = add_text_embeds
        add_time_ids = add_time_ids.tile((self.batch_size * num_images_per_prompt, 1))

        # autoencoder reconstruction
        if self.vae.dtype == ms.float16 and self.vae.config.force_upcast:
            self.upcast_vae()
            x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.get_parameters())).dtype)
            image_rec = self.vae.decode(
                x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator
            )[0]
            # cast back to fp16 if needed
            # TODO: upcast_vae in MS
            self.vae.to(dtype=ms.float16)
        elif self.vae.config.force_upcast:
            x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.get_parameters())).dtype)
            image_rec = self.vae.decode(
                x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator
            )[0]
        else:
            image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]

        image_rec = self.image_processor.postprocess(image_rec, output_type="pil")

        # 5. find zs and xts
        variance_noise_shape = (num_inversion_steps, *x0.shape)

        # intermediate latents
        t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
        xts = mint.zeros(size=variance_noise_shape, dtype=negative_prompt_embeds.dtype)

        for t in reversed(timesteps):
            idx = num_inversion_steps - t_to_idx[int(t)] - 1
            noise = randn_tensor(shape=x0.shape, generator=generator, dtype=x0.dtype)
            xts[idx] = self.scheduler.add_noise(x0, noise, t.unsqueeze(0))
        xts = mint.cat([x0.unsqueeze(0), xts], dim=0)

        # noise maps
        zs = mint.zeros(size=variance_noise_shape, dtype=negative_prompt_embeds.dtype)

        self.scheduler.set_timesteps(len(self.scheduler.timesteps))

        for t in self.progress_bar(timesteps):
            idx = num_inversion_steps - t_to_idx[int(t)] - 1
            # 1. predict noise residual
            xt = xts[idx + 1]

            latent_model_input = mint.cat([xt] * 2) if do_classifier_free_guidance else xt
            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

            added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}

            noise_pred = self.unet(
                latent_model_input,
                t,
                encoder_hidden_states=negative_prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                added_cond_kwargs=added_cond_kwargs,
                return_dict=False,
            )[0]

            # 2. perform guidance
            if do_classifier_free_guidance:
                noise_pred_out = noise_pred.chunk(2)
                noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
                noise_pred = noise_pred_uncond + source_guidance_scale * (noise_pred_text - noise_pred_uncond)

            xtm1 = xts[idx]
            z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, self.eta)
            zs[idx] = z

            # correction to avoid error accumulation
            xts[idx] = xtm1_corrected

        self.init_latents = xts[-1]
        zs = zs.flip((0,))

        if num_zero_noise_steps > 0:
            zs[-num_zero_noise_steps:] = mint.zeros_like(zs[-num_zero_noise_steps:])
        self.zs = zs
        return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec)

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.__call__(denoising_end=None, negative_prompt=None, negative_prompt_2=None, negative_prompt_embeds=None, negative_pooled_prompt_embeds=None, ip_adapter_image=None, output_type='pil', return_dict=False, cross_attention_kwargs=None, guidance_rescale=0.0, crops_coords_top_left=(0, 0), target_size=None, editing_prompt=None, editing_prompt_embeddings=None, editing_pooled_prompt_embeds=None, reverse_editing_direction=False, edit_guidance_scale=5, edit_warmup_steps=0, edit_cooldown_steps=None, edit_threshold=0.9, sem_guidance=None, use_cross_attn_mask=False, use_intersect_mask=False, user_mask=None, attn_store_steps=[], store_averaged_over_steps=True, clip_skip=None, callback_on_step_end=None, callback_on_step_end_tensor_inputs=['latents'], **kwargs)

The call function to the pipeline for editing. The [~pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.invert] method has to be called beforehand. Edits will always be performed for the last inverted image(s).

PARAMETER	DESCRIPTION
denoising_end	When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be completed before it is intentionally prematurely terminated. As a result, the returned sample will still retain a substantial amount of noise as determined by the discrete timesteps selected by the scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image TYPE: `float`, *optional* DEFAULT: None
negative_prompt	The prompt or prompts not to guide the image generation. If not defined, one has to pass negative_prompt_embeds instead. Ignored when not using guidance (i.e., ignored if guidance_scale is less than 1). TYPE: `str` or `List[str]`, *optional* DEFAULT: None
negative_prompt_2	The prompt or prompts not to guide the image generation to be sent to tokenizer_2 and text_encoder_2. If not defined, negative_prompt is used in both text-encoders TYPE: `str` or `List[str]`, *optional* DEFAULT: None
negative_prompt_embeds	Pre-generated negative text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, negative_prompt_embeds will be generated from negative_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
negative_pooled_prompt_embeds	Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, pooled negative_prompt_embeds will be generated from negative_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
ip_adapter_image	(PipelineImageInput, optional): Optional image input to work with IP Adapters. TYPE: Optional[PipelineImageInput] DEFAULT: None
output_type	The output format of the generate image. Choose between PIL: PIL.Image.Image or np.array. TYPE: `str`, *optional*, defaults to `"pil"` DEFAULT: 'pil'
return_dict	Whether or not to return a [~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput] instead of a plain tuple. TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False
callback	A function that will be called every callback_steps steps during inference. The function will be called with the following arguments: callback(step: int, timestep: int, latents: ms.Tensor). TYPE: `Callable`, *optional*
callback_steps	The frequency at which the callback function will be called. If not specified, the callback will be called at every step. TYPE: `int`, *optional*, defaults to 1
cross_attention_kwargs	A kwargs dictionary that if specified is passed along to the AttentionProcessor as defined under self.processor in diffusers.models.attention_processor. TYPE: `dict`, *optional* DEFAULT: None
guidance_rescale	Guidance rescale factor proposed by Common Diffusion Noise Schedules and Sample Steps are Flawed guidance_scale is defined as φ in equation 16. of Common Diffusion Noise Schedules and Sample Steps are Flawed. Guidance rescale factor should fix overexposure when using zero terminal SNR. TYPE: `float`, *optional*, defaults to 0.7 DEFAULT: 0.0
crops_coords_top_left	crops_coords_top_left can be used to generate an image that appears to be "cropped" from the position crops_coords_top_left downwards. Favorable, well-centered images are usually achieved by setting crops_coords_top_left to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of https://huggingface.co/papers/2307.01952. TYPE: `Tuple[int]`, *optional*, defaults to (0, 0 DEFAULT: (0, 0)
target_size	For most cases, target_size should be set to the desired height and width of the generated image. If not specified it will default to (width, height). Part of SDXL's micro-conditioning as explained in section 2.2 of https://huggingface.co/papers/2307.01952. TYPE: `Tuple[int]`, *optional*, defaults to (1024, 1024 DEFAULT: None
editing_prompt	The prompt or prompts to guide the image generation. The image is reconstructed by setting editing_prompt = None. Guidance direction of prompt should be specified via reverse_editing_direction. TYPE: `str` or `List[str]`, *optional* DEFAULT: None
editing_prompt_embeddings	Pre-generated edit text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, editing_prompt_embeddings will be generated from editing_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
editing_pooled_prompt_embeddings	Pre-generated pooled edit text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, editing_prompt_embeddings will be generated from editing_prompt input argument. TYPE: `ms.Tensor`, *optional*
reverse_editing_direction	Whether the corresponding prompt in editing_prompt should be increased or decreased. TYPE: `bool` or `List[bool]`, *optional*, defaults to `False` DEFAULT: False
edit_guidance_scale	Guidance scale for guiding the image generation. If provided as list values should correspond to editing_prompt. edit_guidance_scale is defined as s_e of equation 12 of LEDITS++ Paper. TYPE: `float` or `List[float]`, *optional*, defaults to 5 DEFAULT: 5
edit_warmup_steps	Number of diffusion steps (for each prompt) for which guidance is not applied. TYPE: `float` or `List[float]`, *optional*, defaults to 10 DEFAULT: 0
edit_cooldown_steps	Number of diffusion steps (for each prompt) after which guidance is no longer applied. TYPE: `float` or `List[float]`, *optional*, defaults to `None` DEFAULT: None
edit_threshold	Masking threshold of guidance. Threshold should be proportional to the image region that is modified. 'edit_threshold' is defined as 'λ' of equation 12 of LEDITS++ Paper. TYPE: `float` or `List[float]`, *optional*, defaults to 0.9 DEFAULT: 0.9
sem_guidance	List of pre-generated guidance vectors to be applied at generation. Length of the list has to correspond to num_inference_steps. TYPE: `List[ms.Tensor]`, *optional* DEFAULT: None
use_cross_attn_mask	Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of LEDITS++ paper. TYPE: bool DEFAULT: False
use_intersect_mask	Whether the masking term is calculated as intersection of cross-attention masks and masks derived from the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate are defined as 'M^2' of equation 12 of LEDITS++ paper. TYPE: bool DEFAULT: False
user_mask	User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s implicit masks do not meet user preferences. TYPE: Optional[Tensor] DEFAULT: None
attn_store_steps	Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes. TYPE: Optional[List[int]] DEFAULT: []
store_averaged_over_steps	Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If False, attention maps for each step are stores separately. Just for visualization purposes. TYPE: bool DEFAULT: True
clip_skip	Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. TYPE: `int`, *optional* DEFAULT: None
callback_on_step_end	A function that calls at the end of each denoising steps during the inference. The function is called with the following arguments: callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int, callback_kwargs: Dict). callback_kwargs will include a list of all tensors as specified by callback_on_step_end_tensor_inputs. TYPE: `Callable`, *optional* DEFAULT: None
callback_on_step_end_tensor_inputs	The list of tensor inputs for the callback_on_step_end function. The tensors specified in the list will be passed as callback_kwargs argument. You will only be able to include variables listed in the ._callback_tensor_inputs attribute of your pipeline class. TYPE: `List`, *optional* DEFAULT: ['latents']

RETURNS	DESCRIPTION
	[~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput] or tuple:
	[~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput] if return_dict is True, otherwise a `tuple. When
	returning a tuple, the first element is a list with the generated images.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def __call__(
    self,
    denoising_end: Optional[float] = None,
    negative_prompt: Optional[Union[str, List[str]]] = None,
    negative_prompt_2: Optional[Union[str, List[str]]] = None,
    negative_prompt_embeds: Optional[ms.Tensor] = None,
    negative_pooled_prompt_embeds: Optional[ms.Tensor] = None,
    ip_adapter_image: Optional[PipelineImageInput] = None,
    output_type: Optional[str] = "pil",
    return_dict: bool = False,
    cross_attention_kwargs: Optional[Dict[str, Any]] = None,
    guidance_rescale: float = 0.0,
    crops_coords_top_left: Tuple[int, int] = (0, 0),
    target_size: Optional[Tuple[int, int]] = None,
    editing_prompt: Optional[Union[str, List[str]]] = None,
    editing_prompt_embeddings: Optional[ms.Tensor] = None,
    editing_pooled_prompt_embeds: Optional[ms.Tensor] = None,
    reverse_editing_direction: Optional[Union[bool, List[bool]]] = False,
    edit_guidance_scale: Optional[Union[float, List[float]]] = 5,
    edit_warmup_steps: Optional[Union[int, List[int]]] = 0,
    edit_cooldown_steps: Optional[Union[int, List[int]]] = None,
    edit_threshold: Optional[Union[float, List[float]]] = 0.9,
    sem_guidance: Optional[List[ms.Tensor]] = None,
    use_cross_attn_mask: bool = False,
    use_intersect_mask: bool = False,
    user_mask: Optional[ms.Tensor] = None,
    attn_store_steps: Optional[List[int]] = [],
    store_averaged_over_steps: bool = True,
    clip_skip: Optional[int] = None,
    callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
    callback_on_step_end_tensor_inputs: List[str] = ["latents"],
    **kwargs,
):
    r"""
    The call function to the pipeline for editing. The
    [`~pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.invert`] method has to be called beforehand. Edits
    will always be performed for the last inverted image(s).

    Args:
        denoising_end (`float`, *optional*):
            When specified, determines the fraction (between 0.0 and 1.0) of the total denoising process to be
            completed before it is intentionally prematurely terminated. As a result, the returned sample will
            still retain a substantial amount of noise as determined by the discrete timesteps selected by the
            scheduler. The denoising_end parameter should ideally be utilized when this pipeline forms a part of a
            "Mixture of Denoisers" multi-pipeline setup, as elaborated in [**Refining the Image
        negative_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation. If not defined, one has to pass
            `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
            less than `1`).
        negative_prompt_2 (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
            `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
        negative_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
            argument.
        negative_pooled_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
            input argument.
        ip_adapter_image: (`PipelineImageInput`, *optional*):
            Optional image input to work with IP Adapters.
        output_type (`str`, *optional*, defaults to `"pil"`):
            The output format of the generate image. Choose between
            [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
        return_dict (`bool`, *optional*, defaults to `False`):
            Whether or not to return a [`~pipelines.stable_diffusion_xl.StableDiffusionXLPipelineOutput`] instead
            of a plain tuple.
        callback (`Callable`, *optional*):
            A function that will be called every `callback_steps` steps during inference. The function will be
            called with the following arguments: `callback(step: int, timestep: int, latents: ms.Tensor)`.
        callback_steps (`int`, *optional*, defaults to 1):
            The frequency at which the `callback` function will be called. If not specified, the callback will be
            called at every step.
        cross_attention_kwargs (`dict`, *optional*):
            A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
            `self.processor` in
            [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
        guidance_rescale (`float`, *optional*, defaults to 0.7):
            Guidance rescale factor proposed by [Common Diffusion Noise Schedules and Sample Steps are
            Flawed](https://arxiv.org/pdf/2305.08891.pdf) `guidance_scale` is defined as `φ` in equation 16. of
            [Common Diffusion Noise Schedules and Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf).
            Guidance rescale factor should fix overexposure when using zero terminal SNR.
        crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
            `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
            `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
            `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
            [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
        target_size (`Tuple[int]`, *optional*, defaults to (1024, 1024)):
            For most cases, `target_size` should be set to the desired height and width of the generated image. If
            not specified it will default to `(width, height)`. Part of SDXL's micro-conditioning as explained in
            section 2.2 of [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
        editing_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts to guide the image generation. The image is reconstructed by setting
            `editing_prompt = None`. Guidance direction of prompt should be specified via
            `reverse_editing_direction`.
        editing_prompt_embeddings (`ms.Tensor`, *optional*):
            Pre-generated edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
            If not provided, editing_prompt_embeddings will be generated from `editing_prompt` input argument.
        editing_pooled_prompt_embeddings (`ms.Tensor`, *optional*):
            Pre-generated pooled edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, editing_prompt_embeddings will be generated from `editing_prompt` input
            argument.
        reverse_editing_direction (`bool` or `List[bool]`, *optional*, defaults to `False`):
            Whether the corresponding prompt in `editing_prompt` should be increased or decreased.
        edit_guidance_scale (`float` or `List[float]`, *optional*, defaults to 5):
            Guidance scale for guiding the image generation. If provided as list values should correspond to
            `editing_prompt`. `edit_guidance_scale` is defined as `s_e` of equation 12 of [LEDITS++
            Paper](https://arxiv.org/abs/2301.12247).
        edit_warmup_steps (`float` or `List[float]`, *optional*, defaults to 10):
            Number of diffusion steps (for each prompt) for which guidance is not applied.
        edit_cooldown_steps (`float` or `List[float]`, *optional*, defaults to `None`):
            Number of diffusion steps (for each prompt) after which guidance is no longer applied.
        edit_threshold (`float` or `List[float]`, *optional*, defaults to 0.9):
            Masking threshold of guidance. Threshold should be proportional to the image region that is modified.
            'edit_threshold' is defined as 'λ' of equation 12 of [LEDITS++
            Paper](https://arxiv.org/abs/2301.12247).
        sem_guidance (`List[ms.Tensor]`, *optional*):
            List of pre-generated guidance vectors to be applied at generation. Length of the list has to
            correspond to `num_inference_steps`.
        use_cross_attn_mask:
            Whether cross-attention masks are used. Cross-attention masks are always used when use_intersect_mask
            is set to true. Cross-attention masks are defined as 'M^1' of equation 12 of [LEDITS++
            paper](https://arxiv.org/pdf/2311.16711.pdf).
        use_intersect_mask:
            Whether the masking term is calculated as intersection of cross-attention masks and masks derived from
            the noise estimate. Cross-attention mask are defined as 'M^1' and masks derived from the noise estimate
            are defined as 'M^2' of equation 12 of [LEDITS++ paper](https://arxiv.org/pdf/2311.16711.pdf).
        user_mask:
            User-provided mask for even better control over the editing process. This is helpful when LEDITS++'s
            implicit masks do not meet user preferences.
        attn_store_steps:
            Steps for which the attention maps are stored in the AttentionStore. Just for visualization purposes.
        store_averaged_over_steps:
            Whether the attention maps for the 'attn_store_steps' are stored averaged over the diffusion steps. If
            False, attention maps for each step are stores separately. Just for visualization purposes.
        clip_skip (`int`, *optional*):
            Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
            the output of the pre-final layer will be used for computing the prompt embeddings.
        callback_on_step_end (`Callable`, *optional*):
            A function that calls at the end of each denoising steps during the inference. The function is called
            with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
            callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
            `callback_on_step_end_tensor_inputs`.
        callback_on_step_end_tensor_inputs (`List`, *optional*):
            The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
            will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
            `._callback_tensor_inputs` attribute of your pipeline class.

    Examples:

    Returns:
        [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] or `tuple`:
        [`~pipelines.ledits_pp.LEditsPPDiffusionPipelineOutput`] if `return_dict` is True, otherwise a `tuple. When
        returning a tuple, the first element is a list with the generated images.
    """
    if self.inversion_steps is None:
        raise ValueError(
            "You need to invert an input image first before calling the pipeline. The `invert` method has to be called beforehand. Edits will always be performed for the last inverted image(s)."  # noqa: E501
        )

    eta = self.eta
    num_images_per_prompt = 1
    latents = self.init_latents

    zs = self.zs
    self.scheduler.set_timesteps(len(self.scheduler.timesteps))

    if use_intersect_mask:
        use_cross_attn_mask = True

    if use_cross_attn_mask:
        self.smoothing = LeditsGaussianSmoothing()

    # TODO: Check inputs
    # 1. Check inputs. Raise error if not correct
    # self.check_inputs(
    #    callback_steps,
    #    negative_prompt,
    #    negative_prompt_2,
    #    prompt_embeds,
    #    negative_prompt_embeds,
    #    pooled_prompt_embeds,
    #    negative_pooled_prompt_embeds,
    # )
    self._guidance_rescale = guidance_rescale
    self._clip_skip = clip_skip
    self._cross_attention_kwargs = cross_attention_kwargs
    self._denoising_end = denoising_end

    # 2. Define call parameters
    batch_size = self.batch_size

    if editing_prompt:
        enable_edit_guidance = True
        if isinstance(editing_prompt, str):
            editing_prompt = [editing_prompt]
        self.enabled_editing_prompts = len(editing_prompt)
    elif editing_prompt_embeddings is not None:
        enable_edit_guidance = True
        self.enabled_editing_prompts = editing_prompt_embeddings.shape[0]
    else:
        self.enabled_editing_prompts = 0
        enable_edit_guidance = False

    # 3. Encode input prompt
    text_encoder_lora_scale = (
        cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
    )
    (
        prompt_embeds,
        edit_prompt_embeds,
        negative_pooled_prompt_embeds,
        pooled_edit_embeds,
        num_edit_tokens,
    ) = self.encode_prompt(
        num_images_per_prompt=num_images_per_prompt,
        negative_prompt=negative_prompt,
        negative_prompt_2=negative_prompt_2,
        negative_prompt_embeds=negative_prompt_embeds,
        negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
        lora_scale=text_encoder_lora_scale,
        clip_skip=self.clip_skip,
        enable_edit_guidance=enable_edit_guidance,
        editing_prompt=editing_prompt,
        editing_prompt_embeds=editing_prompt_embeddings,
        editing_pooled_prompt_embeds=editing_pooled_prompt_embeds,
    )

    # 4. Prepare timesteps
    # self.scheduler.set_timesteps(num_inference_steps)

    timesteps = self.inversion_steps
    t_to_idx = {int(v): k for k, v in enumerate(timesteps)}

    if use_cross_attn_mask:
        self.attention_store = LeditsAttentionStore(
            average=store_averaged_over_steps,
            batch_size=batch_size,
            max_size=(latents.shape[-2] / 4.0) * (latents.shape[-1] / 4.0),
            max_resolution=None,
        )
        self.prepare_unet(self.attention_store)
        resolution = latents.shape[-2:]
        att_res = (int(resolution[0] / 4), int(resolution[1] / 4))

    # 5. Prepare latent variables
    latents = self.prepare_latents(latents=latents)

    # 6. Prepare extra step kwargs.
    extra_step_kwargs = self.prepare_extra_step_kwargs(eta)

    if self.text_encoder_2 is None:
        text_encoder_projection_dim = int(negative_pooled_prompt_embeds.shape[-1])
    else:
        text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

    # 7. Prepare added time ids & embeddings
    add_text_embeds = negative_pooled_prompt_embeds
    add_time_ids = self._get_add_time_ids(
        self.size,
        crops_coords_top_left,
        self.size,
        dtype=negative_pooled_prompt_embeds.dtype,
        text_encoder_projection_dim=text_encoder_projection_dim,
    )

    if enable_edit_guidance:
        prompt_embeds = mint.cat([prompt_embeds, edit_prompt_embeds], dim=0)
        add_text_embeds = mint.cat([add_text_embeds, pooled_edit_embeds], dim=0)
        edit_concepts_time_ids = add_time_ids.tile((edit_prompt_embeds.shape[0], 1))
        add_time_ids = mint.cat([add_time_ids, edit_concepts_time_ids], dim=0)
        self.text_cross_attention_maps = [editing_prompt] if isinstance(editing_prompt, str) else editing_prompt

    prompt_embeds = prompt_embeds
    add_text_embeds = add_text_embeds
    add_time_ids = add_time_ids.tile((batch_size * num_images_per_prompt, 1))

    if ip_adapter_image is not None:
        # TODO: fix image encoding
        image_embeds, negative_image_embeds = self.encode_image(ip_adapter_image, num_images_per_prompt)
        if self.do_classifier_free_guidance:
            image_embeds = mint.cat([negative_image_embeds, image_embeds])
            image_embeds = image_embeds

    # 8. Denoising loop
    self.sem_guidance = None
    self.activation_mask = None

    if (
        self.denoising_end is not None
        and isinstance(self.denoising_end, float)
        and self.denoising_end > 0
        and self.denoising_end < 1
    ):
        discrete_timestep_cutoff = int(
            round(
                self.scheduler.config.num_train_timesteps
                - (self.denoising_end * self.scheduler.config.num_train_timesteps)
            )
        )
        num_inference_steps = len(list(filter(lambda ts: ts >= discrete_timestep_cutoff, timesteps)))
        timesteps = timesteps[:num_inference_steps]

    # 9. Optionally get Guidance Scale Embedding
    timestep_cond = None
    if self.unet.config.time_cond_proj_dim is not None:
        guidance_scale_tensor = ms.tensor(self.guidance_scale - 1).tile((batch_size * num_images_per_prompt))
        timestep_cond = self.get_guidance_scale_embedding(
            guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
        ).to(dtype=latents.dtype)

    self._num_timesteps = len(timesteps)
    with self.progress_bar(total=self._num_timesteps) as progress_bar:
        for i, t in enumerate(timesteps):
            # expand the latents if we are doing classifier free guidance
            latent_model_input = mint.cat([latents] * (1 + self.enabled_editing_prompts))
            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
            # predict the noise residual
            added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}
            if ip_adapter_image is not None:
                added_cond_kwargs["image_embeds"] = image_embeds
            noise_pred = self.unet(
                latent_model_input,
                t,
                encoder_hidden_states=prompt_embeds,
                cross_attention_kwargs=cross_attention_kwargs,
                added_cond_kwargs=added_cond_kwargs,
                return_dict=False,
            )[0]

            noise_pred_out = noise_pred.chunk(1 + self.enabled_editing_prompts)  # [b,4, 64, 64]
            noise_pred_uncond = noise_pred_out[0]
            noise_pred_edit_concepts = noise_pred_out[1:]

            noise_guidance_edit = mint.zeros(
                noise_pred_uncond.shape,
                dtype=noise_pred_uncond.dtype,
            )

            if sem_guidance is not None and len(sem_guidance) > i:
                noise_guidance_edit += sem_guidance[i]

            elif enable_edit_guidance:
                if self.activation_mask is None:
                    self.activation_mask = mint.zeros(
                        (len(timesteps), self.enabled_editing_prompts, *noise_pred_edit_concepts[0].shape)
                    )
                if self.sem_guidance is None:
                    self.sem_guidance = mint.zeros((len(timesteps), *noise_pred_uncond.shape))

                # noise_guidance_edit = mint.zeros_like(noise_guidance)
                for c, noise_pred_edit_concept in enumerate(noise_pred_edit_concepts):
                    if isinstance(edit_warmup_steps, list):
                        edit_warmup_steps_c = edit_warmup_steps[c]
                    else:
                        edit_warmup_steps_c = edit_warmup_steps
                    if i < edit_warmup_steps_c:
                        continue

                    if isinstance(edit_guidance_scale, list):
                        edit_guidance_scale_c = edit_guidance_scale[c]
                    else:
                        edit_guidance_scale_c = edit_guidance_scale

                    if isinstance(edit_threshold, list):
                        edit_threshold_c = edit_threshold[c]
                    else:
                        edit_threshold_c = edit_threshold
                    if isinstance(reverse_editing_direction, list):
                        reverse_editing_direction_c = reverse_editing_direction[c]
                    else:
                        reverse_editing_direction_c = reverse_editing_direction

                    if isinstance(edit_cooldown_steps, list):
                        edit_cooldown_steps_c = edit_cooldown_steps[c]
                    elif edit_cooldown_steps is None:
                        edit_cooldown_steps_c = i + 1
                    else:
                        edit_cooldown_steps_c = edit_cooldown_steps

                    if i >= edit_cooldown_steps_c:
                        continue

                    noise_guidance_edit_tmp = noise_pred_edit_concept - noise_pred_uncond

                    if reverse_editing_direction_c:
                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * -1

                    noise_guidance_edit_tmp = noise_guidance_edit_tmp * edit_guidance_scale_c

                    if user_mask is not None:
                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * user_mask

                    if use_cross_attn_mask:
                        out = self.attention_store.aggregate_attention(
                            attention_maps=self.attention_store.step_store,
                            prompts=self.text_cross_attention_maps,
                            res=att_res,
                            from_where=["up", "down"],
                            is_cross=True,
                            select=self.text_cross_attention_maps.index(editing_prompt[c]),
                        )
                        attn_map = out[:, :, :, 1 : 1 + num_edit_tokens[c]]  # 0 -> startoftext

                        # average over all tokens
                        if attn_map.shape[3] != num_edit_tokens[c]:
                            raise ValueError(
                                f"Incorrect shape of attention_map. Expected size {num_edit_tokens[c]}, but found {attn_map.shape[3]}!"
                            )
                        attn_map = mint.sum(attn_map, dim=3)

                        # gaussian_smoothing
                        attn_map = mint.nn.functional.pad(attn_map.unsqueeze(1), (1, 1, 1, 1), mode="reflect")
                        attn_map = self.smoothing(attn_map).squeeze(1)

                        # ops.quantile function expects float32
                        # TODO: ops.quantile is not supported
                        if attn_map.dtype == ms.float32:
                            tmp = ms.tensor(
                                np.quantile(attn_map.flatten(start_dim=1).numpy(), edit_threshold_c, axis=1)
                            )
                        else:
                            tmp = ms.tensor(
                                np.quantile(
                                    attn_map.flatten(start_dim=1).to(ms.float32).numpy(), edit_threshold_c, axis=1
                                )
                            ).to(attn_map.dtype)
                        attn_mask = mint.where(
                            attn_map >= tmp.unsqueeze(1).unsqueeze(1).tile((1, *att_res)), 1.0, 0.0
                        )

                        # resolution must match latent space dimension
                        attn_mask = mint.nn.functional.interpolate(
                            attn_mask.unsqueeze(1),
                            noise_guidance_edit_tmp.shape[-2:],  # 64,64
                        ).tile((1, 4, 1, 1))
                        self.activation_mask[i, c] = ops.stop_gradient(attn_mask)
                        if not use_intersect_mask:
                            noise_guidance_edit_tmp = noise_guidance_edit_tmp * attn_mask

                    if use_intersect_mask:
                        noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                        noise_guidance_edit_tmp_quantile = mint.sum(
                            noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                        )
                        noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile(
                            (1, self.unet.config.in_channels, 1, 1)
                        )

                        # ops.quantile function expects float32
                        # TODO: ops.quantile is not supported
                        if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                            tmp = ms.tensor(
                                np.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                    edit_threshold_c,
                                    axis=2,
                                    keepdims=False,
                                )
                            )
                        else:
                            tmp = ms.tensor(
                                np.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(ms.float32).numpy(),
                                    edit_threshold_c,
                                    axis=2,
                                    keepdims=False,
                                )
                            ).to(noise_guidance_edit_tmp_quantile.dtype)

                        intersect_mask = (
                            mint.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                mint.ones_like(noise_guidance_edit_tmp),
                                mint.zeros_like(noise_guidance_edit_tmp),
                            )
                            * attn_mask
                        )

                        self.activation_mask[i, c] = ops.stop_gradient(intersect_mask)

                        noise_guidance_edit_tmp = noise_guidance_edit_tmp * intersect_mask

                    elif not use_cross_attn_mask:
                        # calculate quantile
                        noise_guidance_edit_tmp_quantile = mint.abs(noise_guidance_edit_tmp)
                        noise_guidance_edit_tmp_quantile = mint.sum(
                            noise_guidance_edit_tmp_quantile, dim=1, keepdim=True
                        )
                        noise_guidance_edit_tmp_quantile = noise_guidance_edit_tmp_quantile.tile((1, 4, 1, 1))

                        # ops.quantile function expects float32
                        # TODO: ops.quantile is not supported
                        if noise_guidance_edit_tmp_quantile.dtype == ms.float32:
                            tmp = ms.tensor(
                                np.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2).numpy(),
                                    edit_threshold_c,
                                    axis=2,
                                    keepdims=False,
                                )
                            )
                        else:
                            tmp = ms.tensor(
                                np.quantile(
                                    noise_guidance_edit_tmp_quantile.flatten(start_dim=2).to(ms.float32).numpy(),
                                    edit_threshold_c,
                                    axis=2,
                                    keepdims=False,
                                )
                            ).to(noise_guidance_edit_tmp_quantile.dtype)

                        self.activation_mask[i, c] = ops.stop_gradient(
                            mint.where(
                                noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                                mint.ones_like(noise_guidance_edit_tmp),
                                mint.zeros_like(noise_guidance_edit_tmp),
                            )
                        )

                        noise_guidance_edit_tmp = mint.where(
                            noise_guidance_edit_tmp_quantile >= tmp[:, :, None, None],
                            noise_guidance_edit_tmp,
                            mint.zeros_like(noise_guidance_edit_tmp),
                        )

                    noise_guidance_edit += noise_guidance_edit_tmp

                self.sem_guidance[i] = ops.stop_gradient(noise_guidance_edit)

            noise_pred = noise_pred_uncond + noise_guidance_edit

            # compute the previous noisy sample x_t -> x_t-1
            if enable_edit_guidance and self.guidance_rescale > 0.0:
                # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
                noise_pred = rescale_noise_cfg(
                    noise_pred,
                    noise_pred_edit_concepts.mean(dim=0, keepdim=False),
                    guidance_rescale=self.guidance_rescale,
                )

            idx = t_to_idx[int(t)]
            latents = self.scheduler.step(
                noise_pred, t, latents, variance_noise=zs[idx], **extra_step_kwargs, return_dict=False
            )[0]

            # step callback
            if use_cross_attn_mask:
                store_step = i in attn_store_steps
                self.attention_store.between_steps(store_step)

            if callback_on_step_end is not None:
                callback_kwargs = {}
                for k in callback_on_step_end_tensor_inputs:
                    callback_kwargs[k] = locals()[k]
                callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)

                latents = callback_outputs.pop("latents", latents)
                prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
                negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
                add_text_embeds = callback_outputs.pop("add_text_embeds", add_text_embeds)
                negative_pooled_prompt_embeds = callback_outputs.pop(
                    "negative_pooled_prompt_embeds", negative_pooled_prompt_embeds
                )
                add_time_ids = callback_outputs.pop("add_time_ids", add_time_ids)
                # negative_add_time_ids = callback_outputs.pop("negative_add_time_ids", negative_add_time_ids)

            # call the callback, if provided
            if i == len(timesteps) - 1 or ((i + 1) > 0 and (i + 1) % self.scheduler.order == 0):
                progress_bar.update()

    if not output_type == "latent":
        # make sure the VAE is in float32 mode, as it overflows in float16
        needs_upcasting = self.vae.dtype == ms.float16 and self.vae.config.force_upcast

        if needs_upcasting:
            self.upcast_vae()
            latents = latents.to(next(iter(self.vae.post_quant_conv.get_parameters())).dtype)

        image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]

        # cast back to fp16 if needed
        if needs_upcasting:
            self.vae.to(dtype=ms.float16)
    else:
        image = latents

    if not output_type == "latent":
        # apply watermark if available
        if self.watermark is not None:
            image = self.watermark.apply_watermark(image)

        image = self.image_processor.postprocess(image, output_type=output_type)

    if not return_dict:
        return (image,)

    return LEditsPPDiffusionPipelineOutput(images=image, nsfw_content_detected=None)

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.disable_vae_slicing()

Disable sliced VAE decoding. If enable_vae_slicing was previously enabled, this method will go back to computing decoding in one step.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def disable_vae_slicing(self):
    r"""
    Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
    computing decoding in one step.
    """
    self.vae.disable_slicing()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.disable_vae_tiling()

Disable tiled VAE decoding. If enable_vae_tiling was previously enabled, this method will go back to computing decoding in one step.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def disable_vae_tiling(self):
    r"""
    Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
    computing decoding in one step.
    """
    self.vae.disable_tiling()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.enable_vae_slicing()

Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def enable_vae_slicing(self):
    r"""
    Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
    compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
    """
    self.vae.enable_slicing()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.enable_vae_tiling()

Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow processing larger images.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def enable_vae_tiling(self):
    r"""
    Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
    compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
    processing larger images.
    """
    self.vae.enable_tiling()

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.encode_prompt(num_images_per_prompt=1, negative_prompt=None, negative_prompt_2=None, negative_prompt_embeds=None, negative_pooled_prompt_embeds=None, lora_scale=None, clip_skip=None, enable_edit_guidance=True, editing_prompt=None, editing_prompt_embeds=None, editing_pooled_prompt_embeds=None)

Encodes the prompt into text encoder hidden states.

PARAMETER	DESCRIPTION
num_images_per_prompt	number of images that should be generated per prompt TYPE: `int` DEFAULT: 1
negative_prompt	The prompt or prompts not to guide the image generation. If not defined, one has to pass negative_prompt_embeds instead. TYPE: `str` or `List[str]`, *optional* DEFAULT: None
negative_prompt_2	The prompt or prompts not to guide the image generation to be sent to tokenizer_2 and text_encoder_2. If not defined, negative_prompt is used in both text-encoders TYPE: `str` or `List[str]`, *optional* DEFAULT: None
negative_prompt_embeds	Pre-generated negative text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, negative_prompt_embeds will be generated from negative_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
negative_pooled_prompt_embeds	Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, pooled negative_prompt_embeds will be generated from negative_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
lora_scale	A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded. TYPE: `float`, *optional* DEFAULT: None
clip_skip	Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that the output of the pre-final layer will be used for computing the prompt embeddings. TYPE: `int`, *optional* DEFAULT: None
enable_edit_guidance	Whether to guide towards an editing prompt or not. TYPE: `bool` DEFAULT: True
editing_prompt	Editing prompt(s) to be encoded. If not defined and 'enable_edit_guidance' is True, one has to pass editing_prompt_embeds instead. TYPE: `str` or `List[str]`, *optional* DEFAULT: None
editing_prompt_embeds	Pre-generated edit text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided and 'enable_edit_guidance' is True, editing_prompt_embeds will be generated from editing_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None
editing_pooled_prompt_embeds	Pre-generated edit pooled text embeddings. Can be used to easily tweak text inputs, e.g. prompt weighting. If not provided, pooled editing_pooled_prompt_embeds will be generated from editing_prompt input argument. TYPE: `ms.Tensor`, *optional* DEFAULT: None

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def encode_prompt(
    self,
    num_images_per_prompt: int = 1,
    negative_prompt: Optional[str] = None,
    negative_prompt_2: Optional[str] = None,
    negative_prompt_embeds: Optional[ms.Tensor] = None,
    negative_pooled_prompt_embeds: Optional[ms.Tensor] = None,
    lora_scale: Optional[float] = None,
    clip_skip: Optional[int] = None,
    enable_edit_guidance: bool = True,
    editing_prompt: Optional[str] = None,
    editing_prompt_embeds: Optional[ms.Tensor] = None,
    editing_pooled_prompt_embeds: Optional[ms.Tensor] = None,
) -> object:
    r"""
    Encodes the prompt into text encoder hidden states.

    Args:
        num_images_per_prompt (`int`):
            number of images that should be generated per prompt
        negative_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation. If not defined, one has to pass
            `negative_prompt_embeds` instead.
        negative_prompt_2 (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
            `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
        negative_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
            argument.
        negative_pooled_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_prompt`
            input argument.
        lora_scale (`float`, *optional*):
            A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        clip_skip (`int`, *optional*):
            Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
            the output of the pre-final layer will be used for computing the prompt embeddings.
        enable_edit_guidance (`bool`):
            Whether to guide towards an editing prompt or not.
        editing_prompt (`str` or `List[str]`, *optional*):
            Editing prompt(s) to be encoded. If not defined and 'enable_edit_guidance' is True, one has to pass
            `editing_prompt_embeds` instead.
        editing_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated edit text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
            If not provided and 'enable_edit_guidance' is True, editing_prompt_embeds will be generated from
            `editing_prompt` input argument.
        editing_pooled_prompt_embeds (`ms.Tensor`, *optional*):
            Pre-generated edit pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
            weighting. If not provided, pooled editing_pooled_prompt_embeds will be generated from `editing_prompt`
            input argument.
    """
    # set lora scale so that monkey patched LoRA
    # function of text encoder can correctly access it
    if lora_scale is not None and isinstance(self, StableDiffusionXLLoraLoaderMixin):
        self._lora_scale = lora_scale

        # dynamically adjust the LoRA scale
        if self.text_encoder is not None:
            scale_lora_layers(self.text_encoder, lora_scale)

        if self.text_encoder_2 is not None:
            scale_lora_layers(self.text_encoder_2, lora_scale)

    batch_size = self.batch_size

    # Define tokenizers and text encoders
    tokenizers = [self.tokenizer, self.tokenizer_2] if self.tokenizer is not None else [self.tokenizer_2]
    text_encoders = (
        [self.text_encoder, self.text_encoder_2] if self.text_encoder is not None else [self.text_encoder_2]
    )
    num_edit_tokens = 0

    # get unconditional embeddings for classifier free guidance
    zero_out_negative_prompt = negative_prompt is None and self.config.force_zeros_for_empty_prompt

    if negative_prompt_embeds is None:
        negative_prompt = negative_prompt or ""
        negative_prompt_2 = negative_prompt_2 or negative_prompt

        # normalize str to list
        negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
        negative_prompt_2 = (
            batch_size * [negative_prompt_2] if isinstance(negative_prompt_2, str) else negative_prompt_2
        )

        uncond_tokens: List[str]

        if batch_size != len(negative_prompt):
            raise ValueError(
                f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but image inversion "
                f" has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
                " the batch size of the input images."
            )
        else:
            uncond_tokens = [negative_prompt, negative_prompt_2]

        negative_prompt_embeds_list = []
        for negative_prompt, tokenizer, text_encoder in zip(uncond_tokens, tokenizers, text_encoders):
            if isinstance(self, TextualInversionLoaderMixin):
                negative_prompt = self.maybe_convert_prompt(negative_prompt, tokenizer)

            uncond_input = tokenizer(
                negative_prompt,
                padding="max_length",
                max_length=tokenizer.model_max_length,
                truncation=True,
                return_tensors="np",
            )

            negative_prompt_embeds = text_encoder(
                ms.tensor(uncond_input.input_ids),
                output_hidden_states=True,
            )
            # We are only ALWAYS interested in the pooled output of the final text encoder
            negative_pooled_prompt_embeds = negative_prompt_embeds[0]
            negative_prompt_embeds = negative_prompt_embeds[2][-2]

            negative_prompt_embeds_list.append(negative_prompt_embeds)

        negative_prompt_embeds = mint.concat(negative_prompt_embeds_list, dim=-1)

        if zero_out_negative_prompt:
            negative_prompt_embeds = mint.zeros_like(negative_prompt_embeds)
            negative_pooled_prompt_embeds = mint.zeros_like(negative_pooled_prompt_embeds)

    if enable_edit_guidance and editing_prompt_embeds is None:
        editing_prompt_2 = editing_prompt

        editing_prompts = [editing_prompt, editing_prompt_2]
        edit_prompt_embeds_list = []

        for editing_prompt, tokenizer, text_encoder in zip(editing_prompts, tokenizers, text_encoders):
            if isinstance(self, TextualInversionLoaderMixin):
                editing_prompt = self.maybe_convert_prompt(editing_prompt, tokenizer)

            max_length = negative_prompt_embeds.shape[1]
            edit_concepts_input = tokenizer(
                # [x for item in editing_prompt for x in repeat(item, batch_size)],
                editing_prompt,
                padding="max_length",
                max_length=max_length,
                truncation=True,
                return_tensors="np",
                return_length=True,
            )
            num_edit_tokens = edit_concepts_input.length - 2

            edit_concepts_embeds = text_encoder(
                ms.tensor(edit_concepts_input.input_ids),
                output_hidden_states=True,
            )
            # We are only ALWAYS interested in the pooled output of the final text encoder
            editing_pooled_prompt_embeds = edit_concepts_embeds[0]
            if clip_skip is None:
                edit_concepts_embeds = edit_concepts_embeds[2][-2]
            else:
                # "2" because SDXL always indexes from the penultimate layer.
                edit_concepts_embeds = edit_concepts_embeds[-2][-(clip_skip + 2)]

            edit_prompt_embeds_list.append(edit_concepts_embeds)

        edit_concepts_embeds = mint.concat(edit_prompt_embeds_list, dim=-1)
    elif not enable_edit_guidance:
        edit_concepts_embeds = None
        editing_pooled_prompt_embeds = None

    negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype)
    bs_embed, seq_len, _ = negative_prompt_embeds.shape
    # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
    seq_len = negative_prompt_embeds.shape[1]
    negative_prompt_embeds = negative_prompt_embeds.to(dtype=self.text_encoder_2.dtype)
    negative_prompt_embeds = negative_prompt_embeds.tile((1, num_images_per_prompt, 1))
    negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

    if enable_edit_guidance:
        bs_embed_edit, seq_len, _ = edit_concepts_embeds.shape
        edit_concepts_embeds = edit_concepts_embeds.to(dtype=self.text_encoder_2.dtype)
        edit_concepts_embeds = edit_concepts_embeds.tile((1, num_images_per_prompt, 1))
        edit_concepts_embeds = edit_concepts_embeds.view(bs_embed_edit * num_images_per_prompt, seq_len, -1)

    negative_pooled_prompt_embeds = negative_pooled_prompt_embeds.tile((1, num_images_per_prompt)).view(
        bs_embed * num_images_per_prompt, -1
    )

    if enable_edit_guidance:
        editing_pooled_prompt_embeds = editing_pooled_prompt_embeds.tile((1, num_images_per_prompt)).view(
            bs_embed_edit * num_images_per_prompt, -1
        )

    if self.text_encoder is not None:
        if isinstance(self, StableDiffusionXLLoraLoaderMixin):
            # Retrieve the original scale by scaling back the LoRA layers
            unscale_lora_layers(self.text_encoder, lora_scale)

    if self.text_encoder_2 is not None:
        if isinstance(self, StableDiffusionXLLoraLoaderMixin):
            # Retrieve the original scale by scaling back the LoRA layers
            unscale_lora_layers(self.text_encoder_2, lora_scale)

    return (
        negative_prompt_embeds,
        edit_concepts_embeds,
        negative_pooled_prompt_embeds,
        editing_pooled_prompt_embeds,
        num_edit_tokens,
    )

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.get_guidance_scale_embedding(w, embedding_dim=512, dtype=ms.float32)

See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

PARAMETER	DESCRIPTION
w	Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings. TYPE: `ms.Tensor`
embedding_dim	Dimension of the embeddings to generate. TYPE: `int`, *optional*, defaults to 512 DEFAULT: 512
dtype	Data type of the generated embeddings. TYPE: `ms.Type`, *optional*, defaults to `ms.float32` DEFAULT: float32

RETURNS	DESCRIPTION
Tensor	ms.Tensor: Embedding vectors with shape (len(w), embedding_dim).

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def get_guidance_scale_embedding(
    self, w: ms.Tensor, embedding_dim: int = 512, dtype: ms.Type = ms.float32
) -> ms.Tensor:
    """
    See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298

    Args:
        w (`ms.Tensor`):
            Generate embedding vectors with a specified guidance scale to subsequently enrich timestep embeddings.
        embedding_dim (`int`, *optional*, defaults to 512):
            Dimension of the embeddings to generate.
        dtype (`ms.Type`, *optional*, defaults to `ms.float32`):
            Data type of the generated embeddings.

    Returns:
        `ms.Tensor`: Embedding vectors with shape `(len(w), embedding_dim)`.
    """
    assert len(w.shape) == 1
    w = w * 1000.0

    half_dim = embedding_dim // 2
    emb = mint.log(ms.tensor(10000.0)) / (half_dim - 1)
    emb = mint.exp(mint.arange(half_dim, dtype=dtype) * -emb)
    emb = w.to(dtype)[:, None] * emb[None, :]
    emb = mint.cat([mint.sin(emb), mint.cos(emb)], dim=1)
    if embedding_dim % 2 == 1:  # zero pad
        emb = mint.nn.functional.pad(emb, (0, 1))
    assert emb.shape == (w.shape[0], embedding_dim)
    return emb

mindone.diffusers.pipelines.ledits_pp.LEditsPPPipelineStableDiffusionXL.invert(image, source_prompt='', source_guidance_scale=3.5, negative_prompt=None, negative_prompt_2=None, num_inversion_steps=50, skip=0.15, generator=None, crops_coords_top_left=(0, 0), num_zero_noise_steps=3, cross_attention_kwargs=None, height=None, width=None, resize_mode='default', crops_coords=None)

The function to the pipeline for image inversion as described by the LEDITS++ Paper. If the scheduler is set to [~schedulers.DDIMScheduler] the inversion proposed by edit-friendly DPDM will be performed instead.

PARAMETER	DESCRIPTION
image	Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect ratio. TYPE: `PipelineImageInput`
source_prompt	Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled if the source_prompt is "". TYPE: `str`, defaults to `""` DEFAULT: ''
source_guidance_scale	Strength of guidance during inversion. TYPE: `float`, defaults to `3.5` DEFAULT: 3.5
negative_prompt	The prompt or prompts not to guide the image generation. If not defined, one has to pass negative_prompt_embeds instead. Ignored when not using guidance (i.e., ignored if guidance_scale is less than 1). TYPE: `str` or `List[str]`, *optional* DEFAULT: None
negative_prompt_2	The prompt or prompts not to guide the image generation to be sent to tokenizer_2 and text_encoder_2. If not defined, negative_prompt is used in both text-encoders TYPE: `str` or `List[str]`, *optional* DEFAULT: None
num_inversion_steps	Number of total performed inversion steps after discarding the initial skip steps. TYPE: `int`, defaults to `50` DEFAULT: 50
skip	Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values will lead to stronger changes to the input image. skip has to be between 0 and 1. TYPE: `float`, defaults to `0.15` DEFAULT: 0.15
generator	A np.random.Generator to make inversion deterministic. TYPE: `np.random.Generator`, *optional* DEFAULT: None
crops_coords_top_left	crops_coords_top_left can be used to generate an image that appears to be "cropped" from the position crops_coords_top_left downwards. Favorable, well-centered images are usually achieved by setting crops_coords_top_left to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of https://huggingface.co/papers/2307.01952. TYPE: `Tuple[int]`, *optional*, defaults to (0, 0 DEFAULT: (0, 0)
num_zero_noise_steps	Number of final diffusion steps that will not renoise the current image. If no steps are set to zero SD-XL in combination with [DPMSolverMultistepScheduler] will produce noise artifacts. TYPE: `int`, defaults to `3` DEFAULT: 3
cross_attention_kwargs	A kwargs dictionary that if specified is passed along to the AttentionProcessor as defined under self.processor in diffusers.models.attention_processor. TYPE: `dict`, *optional* DEFAULT: None

RETURNS	DESCRIPTION
	[~pipelines.ledits_pp.LEditsPPInversionPipelineOutput]: Output will contain the resized input image(s)
	and respective VAE reconstruction(s).

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_leditspp_stable_diffusion_xl.py

def invert(
    self,
    image: PipelineImageInput,
    source_prompt: str = "",
    source_guidance_scale=3.5,
    negative_prompt: str = None,
    negative_prompt_2: str = None,
    num_inversion_steps: int = 50,
    skip: float = 0.15,
    generator: Optional[np.random.Generator] = None,
    crops_coords_top_left: Tuple[int, int] = (0, 0),
    num_zero_noise_steps: int = 3,
    cross_attention_kwargs: Optional[Dict[str, Any]] = None,
    height: Optional[int] = None,
    width: Optional[int] = None,
    resize_mode: Optional[str] = "default",
    crops_coords: Optional[Tuple[int, int, int, int]] = None,
):
    r"""
    The function to the pipeline for image inversion as described by the [LEDITS++
    Paper](https://arxiv.org/abs/2301.12247). If the scheduler is set to [`~schedulers.DDIMScheduler`] the
    inversion proposed by [edit-friendly DPDM](https://arxiv.org/abs/2304.06140) will be performed instead.

    Args:
        image (`PipelineImageInput`):
            Input for the image(s) that are to be edited. Multiple input images have to default to the same aspect
            ratio.
        source_prompt (`str`, defaults to `""`):
            Prompt describing the input image that will be used for guidance during inversion. Guidance is disabled
            if the `source_prompt` is `""`.
        source_guidance_scale (`float`, defaults to `3.5`):
            Strength of guidance during inversion.
        negative_prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation. If not defined, one has to pass
            `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `guidance_scale` is
            less than `1`).
        negative_prompt_2 (`str` or `List[str]`, *optional*):
            The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
            `text_encoder_2`. If not defined, `negative_prompt` is used in both text-encoders
        num_inversion_steps (`int`, defaults to `50`):
            Number of total performed inversion steps after discarding the initial `skip` steps.
        skip (`float`, defaults to `0.15`):
            Portion of initial steps that will be ignored for inversion and subsequent generation. Lower values
            will lead to stronger changes to the input image. `skip` has to be between `0` and `1`.
        generator (`np.random.Generator`, *optional*):
            A [`np.random.Generator`](https://numpy.org/doc/stable/reference/random/generator.html) to make inversion
            deterministic.
        crops_coords_top_left (`Tuple[int]`, *optional*, defaults to (0, 0)):
            `crops_coords_top_left` can be used to generate an image that appears to be "cropped" from the position
            `crops_coords_top_left` downwards. Favorable, well-centered images are usually achieved by setting
            `crops_coords_top_left` to (0, 0). Part of SDXL's micro-conditioning as explained in section 2.2 of
            [https://huggingface.co/papers/2307.01952](https://huggingface.co/papers/2307.01952).
        num_zero_noise_steps (`int`, defaults to `3`):
            Number of final diffusion steps that will not renoise the current image. If no steps are set to zero
            SD-XL in combination with [`DPMSolverMultistepScheduler`] will produce noise artifacts.
        cross_attention_kwargs (`dict`, *optional*):
            A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
            `self.processor` in
            [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).

    Returns:
        [`~pipelines.ledits_pp.LEditsPPInversionPipelineOutput`]: Output will contain the resized input image(s)
        and respective VAE reconstruction(s).
    """
    if height is not None and height % 32 != 0 or width is not None and width % 32 != 0:
        raise ValueError("height and width must be a factor of 32.")

    # Reset attn processor, we do not want to store attn maps during inversion
    self.unet.set_attn_processor(AttnProcessor())

    self.eta = 1.0

    self.scheduler.config.timestep_spacing = "leading"
    self.scheduler.set_timesteps(int(num_inversion_steps * (1 + skip)))
    self.inversion_steps = self.scheduler.timesteps[-num_inversion_steps:]
    timesteps = self.inversion_steps

    num_images_per_prompt = 1

    # 0. Ensure that only uncond embedding is used if prompt = ""
    if source_prompt == "":
        # noise pred should only be noise_pred_uncond
        source_guidance_scale = 0.0
        do_classifier_free_guidance = False
    else:
        do_classifier_free_guidance = source_guidance_scale > 1.0

    # 1. prepare image
    x0, resized = self.encode_image(
        image,
        dtype=self.text_encoder_2.dtype,
        height=height,
        width=width,
        resize_mode=resize_mode,
        crops_coords=crops_coords,
    )
    width = x0.shape[2] * self.vae_scale_factor
    height = x0.shape[3] * self.vae_scale_factor
    self.size = (height, width)

    self.batch_size = x0.shape[0]

    # 2. get embeddings
    text_encoder_lora_scale = (
        cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
    )

    if isinstance(source_prompt, str):
        source_prompt = [source_prompt] * self.batch_size

    (
        negative_prompt_embeds,
        prompt_embeds,
        negative_pooled_prompt_embeds,
        edit_pooled_prompt_embeds,
        _,
    ) = self.encode_prompt(
        num_images_per_prompt=num_images_per_prompt,
        negative_prompt=negative_prompt,
        negative_prompt_2=negative_prompt_2,
        editing_prompt=source_prompt,
        lora_scale=text_encoder_lora_scale,
        enable_edit_guidance=do_classifier_free_guidance,
    )
    if self.text_encoder_2 is None:
        text_encoder_projection_dim = int(negative_pooled_prompt_embeds.shape[-1])
    else:
        text_encoder_projection_dim = self.text_encoder_2.config.projection_dim

    # 3. Prepare added time ids & embeddings
    add_text_embeds = negative_pooled_prompt_embeds
    add_time_ids = self._get_add_time_ids(
        self.size,
        crops_coords_top_left,
        self.size,
        dtype=negative_prompt_embeds.dtype,
        text_encoder_projection_dim=text_encoder_projection_dim,
    )

    if do_classifier_free_guidance:
        negative_prompt_embeds = mint.cat([negative_prompt_embeds, prompt_embeds], dim=0)
        add_text_embeds = mint.cat([add_text_embeds, edit_pooled_prompt_embeds], dim=0)
        add_time_ids = mint.cat([add_time_ids, add_time_ids], dim=0)

    negative_prompt_embeds = negative_prompt_embeds

    add_text_embeds = add_text_embeds
    add_time_ids = add_time_ids.tile((self.batch_size * num_images_per_prompt, 1))

    # autoencoder reconstruction
    if self.vae.dtype == ms.float16 and self.vae.config.force_upcast:
        self.upcast_vae()
        x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.get_parameters())).dtype)
        image_rec = self.vae.decode(
            x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator
        )[0]
        # cast back to fp16 if needed
        # TODO: upcast_vae in MS
        self.vae.to(dtype=ms.float16)
    elif self.vae.config.force_upcast:
        x0_tmp = x0.to(next(iter(self.vae.post_quant_conv.get_parameters())).dtype)
        image_rec = self.vae.decode(
            x0_tmp / self.vae.config.scaling_factor, return_dict=False, generator=generator
        )[0]
    else:
        image_rec = self.vae.decode(x0 / self.vae.config.scaling_factor, return_dict=False, generator=generator)[0]

    image_rec = self.image_processor.postprocess(image_rec, output_type="pil")

    # 5. find zs and xts
    variance_noise_shape = (num_inversion_steps, *x0.shape)

    # intermediate latents
    t_to_idx = {int(v): k for k, v in enumerate(timesteps)}
    xts = mint.zeros(size=variance_noise_shape, dtype=negative_prompt_embeds.dtype)

    for t in reversed(timesteps):
        idx = num_inversion_steps - t_to_idx[int(t)] - 1
        noise = randn_tensor(shape=x0.shape, generator=generator, dtype=x0.dtype)
        xts[idx] = self.scheduler.add_noise(x0, noise, t.unsqueeze(0))
    xts = mint.cat([x0.unsqueeze(0), xts], dim=0)

    # noise maps
    zs = mint.zeros(size=variance_noise_shape, dtype=negative_prompt_embeds.dtype)

    self.scheduler.set_timesteps(len(self.scheduler.timesteps))

    for t in self.progress_bar(timesteps):
        idx = num_inversion_steps - t_to_idx[int(t)] - 1
        # 1. predict noise residual
        xt = xts[idx + 1]

        latent_model_input = mint.cat([xt] * 2) if do_classifier_free_guidance else xt
        latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)

        added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}

        noise_pred = self.unet(
            latent_model_input,
            t,
            encoder_hidden_states=negative_prompt_embeds,
            cross_attention_kwargs=cross_attention_kwargs,
            added_cond_kwargs=added_cond_kwargs,
            return_dict=False,
        )[0]

        # 2. perform guidance
        if do_classifier_free_guidance:
            noise_pred_out = noise_pred.chunk(2)
            noise_pred_uncond, noise_pred_text = noise_pred_out[0], noise_pred_out[1]
            noise_pred = noise_pred_uncond + source_guidance_scale * (noise_pred_text - noise_pred_uncond)

        xtm1 = xts[idx]
        z, xtm1_corrected = compute_noise(self.scheduler, xtm1, xt, t, noise_pred, self.eta)
        zs[idx] = z

        # correction to avoid error accumulation
        xts[idx] = xtm1_corrected

    self.init_latents = xts[-1]
    zs = zs.flip((0,))

    if num_zero_noise_steps > 0:
        zs[-num_zero_noise_steps:] = mint.zeros_like(zs[-num_zero_noise_steps:])
    self.zs = zs
    return LEditsPPInversionPipelineOutput(images=resized, vae_reconstruction_images=image_rec)

mindone.diffusers.pipelines.ledits_pp.pipeline_output.LEditsPPDiffusionPipelineOutput dataclass

Bases: BaseOutput

Output class for LEdits++ Diffusion pipelines.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_output.py

@dataclass
class LEditsPPDiffusionPipelineOutput(BaseOutput):
    """
    Output class for LEdits++ Diffusion pipelines.

    Args:
        images (`List[PIL.Image.Image]` or `np.ndarray`)
            List of denoised PIL images of length `batch_size` or NumPy array of shape `(batch_size, height, width,
            num_channels)`.
        nsfw_content_detected (`List[bool]`)
            List indicating whether the corresponding generated image contains “not-safe-for-work” (nsfw) content or
            `None` if safety checking could not be performed.
    """

    images: Union[List[PIL.Image.Image], np.ndarray]
    nsfw_content_detected: Optional[List[bool]]

mindone.diffusers.pipelines.ledits_pp.pipeline_output.LEditsPPInversionPipelineOutput dataclass

Bases: BaseOutput

Output class for LEdits++ Diffusion pipelines.

Source code in mindone/diffusers/pipelines/ledits_pp/pipeline_output.py

@dataclass
class LEditsPPInversionPipelineOutput(BaseOutput):
    """
    Output class for LEdits++ Diffusion pipelines.

    Args:
        input_images (`List[PIL.Image.Image]` or `np.ndarray`)
            List of the cropped and resized input images as PIL images of length `batch_size` or NumPy array of shape `
            (batch_size, height, width, num_channels)`.
        vae_reconstruction_images (`List[PIL.Image.Image]` or `np.ndarray`)
            List of VAE reconstruction of all input images as PIL images of length `batch_size` or NumPy array of shape
            ` (batch_size, height, width, num_channels)`.
    """

    images: Union[List[PIL.Image.Image], np.ndarray]
    vae_reconstruction_images: Union[List[PIL.Image.Image], np.ndarray]