Flux2

Flux.2 is the recent series of image generation models from Black Forest Labs, preceded by the Flux.1 series. It is an entirely new model with a new architecture and pre-training done from scratch!

Original model checkpoints for Flux can be found here. Original inference code can be found here.

Tip

Flux2 can be quite expensive to run on consumer hardware devices. However, you can perform a suite of optimizations to run it faster and in a more memory-friendly manner. Check out this section for more details. Additionally, Flux can benefit from quantization for memory efficiency with a trade-off in inference latency. Refer to this blog post to learn more.

Caption upsampling

Flux.2 can potentially generate better better outputs with better prompts. We can "upsample" an input prompt by setting the caption_upsample_temperature argument in the pipeline call arguments. The official implementation recommends this value to be 0.15.

mindone.diffusers.Flux2Pipeline

Bases: DiffusionPipeline, Flux2LoraLoaderMixin

The Flux2 pipeline for text-to-image generation.

Reference: https://bfl.ai/blog/flux-2

PARAMETER	DESCRIPTION
transformer	Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. TYPE: [`Flux2Transformer2DModel`]
scheduler	A scheduler to be used in combination with transformer to denoise the encoded image latents. TYPE: [`FlowMatchEulerDiscreteScheduler`]
vae	Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. TYPE: [`AutoencoderKLFlux2`]
text_encoder	Mistral3ForConditionalGeneration TYPE: [`Mistral3ForConditionalGeneration`]
tokenizer	Tokenizer of class PixtralProcessor. TYPE: `AutoProcessor`

Source code in mindone/diffusers/pipelines/flux2/pipeline_flux2.py

class Flux2Pipeline(DiffusionPipeline, Flux2LoraLoaderMixin):
    r"""
    The Flux2 pipeline for text-to-image generation.

    Reference: [https://bfl.ai/blog/flux-2](https://bfl.ai/blog/flux-2)

    Args:
        transformer ([`Flux2Transformer2DModel`]):
            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKLFlux2`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`Mistral3ForConditionalGeneration`]):
            [Mistral3ForConditionalGeneration](https://huggingface.co/docs/transformers/en/model_doc/mistral3#transformers.Mistral3ForConditionalGeneration)
        tokenizer (`AutoProcessor`):
            Tokenizer of class
            [PixtralProcessor](https://huggingface.co/docs/transformers/en/model_doc/pixtral#transformers.PixtralProcessor).
    """

    model_cpu_offload_seq = "text_encoder->transformer->vae"
    _callback_tensor_inputs = ["latents", "prompt_embeds"]

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKLFlux2,
        text_encoder: Mistral3ForConditionalGeneration,
        tokenizer: AutoProcessor,
        transformer: Flux2Transformer2DModel,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            tokenizer=tokenizer,
            scheduler=scheduler,
            transformer=transformer,
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1) if getattr(self, "vae", None) else 8
        # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
        self.image_processor = Flux2ImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
        self.tokenizer_max_length = 512
        self.default_sample_size = 128

        self.system_message = SYSTEM_MESSAGE
        self.system_message_upsampling_t2i = SYSTEM_MESSAGE_UPSAMPLING_T2I
        self.system_message_upsampling_i2i = SYSTEM_MESSAGE_UPSAMPLING_I2I
        self.upsampling_max_image_size = UPSAMPLING_MAX_IMAGE_SIZE

    @staticmethod
    def _get_mistral_3_small_prompt_embeds(
        text_encoder: Mistral3ForConditionalGeneration,
        tokenizer: AutoProcessor,
        prompt: Union[str, List[str]],
        dtype: Optional[ms.Type] = None,
        max_sequence_length: int = 512,
        system_message: str = SYSTEM_MESSAGE,
        hidden_states_layers: List[int] = (10, 20, 30),
    ):
        dtype = text_encoder.dtype if dtype is None else dtype

        prompt = [prompt] if isinstance(prompt, str) else prompt

        # Format input messages
        messages_batch = format_input(prompts=prompt, system_message=system_message)

        # Process all messages at once
        inputs = tokenizer.apply_chat_template(
            messages_batch,
            add_generation_prompt=False,
            tokenize=True,
            return_dict=True,
            return_tensors="np",
            padding="max_length",
            truncation=True,
            max_length=max_sequence_length,
        )

        # Move to device
        input_ids = ms.tensor(inputs["input_ids"])
        attention_mask = ms.tensor(inputs["attention_mask"])

        # Forward pass through the model
        output = text_encoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            output_hidden_states=True,
            use_cache=False,
        )

        # Only use outputs from intermediate layers and stack them
        out = mint.stack([output.hidden_states[k] for k in hidden_states_layers], dim=1)
        out = out.to(dtype=dtype)

        batch_size, num_channels, seq_len, hidden_dim = out.shape
        prompt_embeds = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, num_channels * hidden_dim)

        return prompt_embeds

    @staticmethod
    def _prepare_text_ids(
        x: ms.Tensor,  # (B, L, D) or (L, D)
        t_coord: Optional[ms.Tensor] = None,
    ):
        B, L, _ = x.shape
        out_ids = []

        for i in range(B):
            t = mint.arange(1) if t_coord is None else t_coord[i]
            h = mint.arange(1)
            w = mint.arange(1)
            l = mint.arange(L)  # noqa

            coords = cartesian_prod(t, h, w, l)
            out_ids.append(coords)

        return mint.stack(out_ids)

    @staticmethod
    def _prepare_latent_ids(
        latents: ms.Tensor,  # (B, C, H, W)
    ):
        r"""
        Generates 4D position coordinates (T, H, W, L) for latent tensors.

        Args:
            latents (ms.Tensor):
                Latent tensor of shape (B, C, H, W)

        Returns:
            ms.Tensor:
                Position IDs tensor of shape (B, H*W, 4) All batches share the same coordinate structure: T=0,
                H=[0..H-1], W=[0..W-1], L=0
        """

        batch_size, _, height, width = latents.shape

        t = mint.arange(1)  # [0] - time dimension
        h = mint.arange(height)
        w = mint.arange(width)
        # [0] - layer dimension
        l = mint.arange(1)  # noqa

        # Create position IDs: (H*W, 4)
        latent_ids = cartesian_prod(t, h, w, l)

        # Expand to batch: (B, H*W, 4)
        latent_ids = latent_ids.unsqueeze(0).expand((batch_size, -1, -1))

        return latent_ids

    @staticmethod
    def _prepare_image_ids(
        image_latents: List[ms.Tensor],  # [(1, C, H, W), (1, C, H, W), ...]
        scale: int = 10,
    ):
        r"""
        Generates 4D time-space coordinates (T, H, W, L) for a sequence of image latents.

        This function creates a unique coordinate for every pixel/patch across all input latent with different
        dimensions.

        Args:
            image_latents (List[ms.Tensor]):
                A list of image latent feature tensors, typically of shape (C, H, W).
            scale (int, optional):
                A factor used to define the time separation (T-coordinate) between latents. T-coordinate for the i-th
                latent is: 'scale + scale * i'. Defaults to 10.

        Returns:
            ms.Tensor:
                The combined coordinate tensor. Shape: (1, N_total, 4) Where N_total is the sum of (H * W) for all
                input latents.

        Coordinate Components (Dimension 4):
            - T (Time): The unique index indicating which latent image the coordinate belongs to.
            - H (Height): The row index within that latent image.
            - W (Width): The column index within that latent image.
            - L (Seq. Length): A sequence length dimension, which is always fixed at 0 (size 1)
        """

        if not isinstance(image_latents, list):
            raise ValueError(f"Expected `image_latents` to be a list, got {type(image_latents)}.")

        # create time offset for each reference image
        t_coords = [scale + scale * t for t in mint.arange(0, len(image_latents))]
        t_coords = [t.view(-1) for t in t_coords]

        image_latent_ids = []
        for x, t in zip(image_latents, t_coords):
            x = x.squeeze(0)
            _, height, width = x.shape

            x_ids = cartesian_prod(t, mint.arange(height), mint.arange(width), mint.arange(1))
            image_latent_ids.append(x_ids)

        image_latent_ids = mint.cat(image_latent_ids, dim=0)
        image_latent_ids = image_latent_ids.unsqueeze(0)

        return image_latent_ids

    @staticmethod
    def _patchify_latents(latents):
        batch_size, num_channels_latents, height, width = latents.shape
        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
        latents = latents.permute(0, 1, 3, 5, 2, 4)
        latents = latents.reshape(batch_size, num_channels_latents * 4, height // 2, width // 2)
        return latents

    @staticmethod
    def _unpatchify_latents(latents):
        batch_size, num_channels_latents, height, width = latents.shape
        latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), 2, 2, height, width)
        latents = latents.permute(0, 1, 4, 2, 5, 3)
        latents = latents.reshape(batch_size, num_channels_latents // (2 * 2), height * 2, width * 2)
        return latents

    @staticmethod
    def _pack_latents(latents):
        """
        pack latents: (batch_size, num_channels, height, width) -> (batch_size, height * width, num_channels)
        """

        batch_size, num_channels, height, width = latents.shape
        latents = latents.reshape(batch_size, num_channels, height * width).permute(0, 2, 1)

        return latents

    @staticmethod
    def _unpack_latents_with_ids(x: ms.Tensor, x_ids: ms.Tensor) -> list[ms.Tensor]:
        """
        using position ids to scatter tokens into place
        """
        x_list = []
        for data, pos in zip(x, x_ids):
            _, ch = data.shape  # noqa: F841
            h_ids = pos[:, 1].to(ms.int64)
            w_ids = pos[:, 2].to(ms.int64)

            h = mint.max(h_ids) + 1
            w = mint.max(w_ids) + 1

            flat_ids = h_ids * w + w_ids

            out = mint.zeros((h * w, ch), dtype=data.dtype)
            out.scatter_(0, flat_ids.unsqueeze(1).expand((-1, ch)), data)

            # reshape from (H * W, C) to (H, W, C) and permute to (C, H, W)

            out = out.view(h, w, ch).permute(2, 0, 1)
            x_list.append(out)

        return mint.stack(x_list, dim=0)

    def upsample_prompt(
        self,
        prompt: Union[str, List[str]],
        images: Union[List[PIL.Image.Image], List[List[PIL.Image.Image]]] = None,
        temperature: float = 0.15,
    ) -> List[str]:
        prompt = [prompt] if isinstance(prompt, str) else prompt

        # Set system message based on whether images are provided
        if images is None or len(images) == 0 or images[0] is None:
            system_message = SYSTEM_MESSAGE_UPSAMPLING_T2I
        else:
            system_message = SYSTEM_MESSAGE_UPSAMPLING_I2I

        # Validate and process the input images
        if images:
            images = _validate_and_process_images(images, self.image_processor, self.upsampling_max_image_size)

        # Format input messages
        messages_batch = format_input(prompts=prompt, system_message=system_message, images=images)

        # Process all messages at once
        # with image processing a too short max length can throw an error in here.
        inputs = self.tokenizer.apply_chat_template(
            messages_batch,
            add_generation_prompt=True,
            tokenize=True,
            return_dict=True,
            return_tensors="np",
            padding="max_length",
            truncation=True,
            max_length=2048,
        )

        # Move to device
        inputs["input_ids"] = ms.tensor(inputs["input_ids"])
        inputs["attention_mask"] = ms.tensor(inputs["attention_mask"])

        if "pixel_values" in inputs:
            inputs["pixel_values"] = ms.tensor(inputs["pixel_values"]).to(self.text_encoder.dtype)

        # Generate text using the model's generate method
        generated_ids = self.text_encoder.generate(
            **inputs,
            max_new_tokens=512,
            do_sample=True,
            temperature=temperature,
            use_cache=True,
        )

        # Decode only the newly generated tokens (skip input tokens)
        # Extract only the generated portion
        input_length = inputs["input_ids"].shape[1]
        generated_tokens = generated_ids[:, input_length:]

        upsampled_prompt = self.tokenizer.tokenizer.batch_decode(
            generated_tokens, skip_special_tokens=True, clean_up_tokenization_spaces=True
        )
        return upsampled_prompt

    def encode_prompt(
        self,
        prompt: Union[str, List[str]],
        num_images_per_prompt: int = 1,
        prompt_embeds: Optional[ms.Tensor] = None,
        max_sequence_length: int = 512,
        text_encoder_out_layers: Tuple[int] = (10, 20, 30),
    ):
        if prompt is None:
            prompt = ""

        prompt = [prompt] if isinstance(prompt, str) else prompt

        if prompt_embeds is None:
            prompt_embeds = self._get_mistral_3_small_prompt_embeds(
                text_encoder=self.text_encoder,
                tokenizer=self.tokenizer,
                prompt=prompt,
                max_sequence_length=max_sequence_length,
                system_message=self.system_message,
                hidden_states_layers=text_encoder_out_layers,
            )

        batch_size, seq_len, _ = prompt_embeds.shape
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        text_ids = self._prepare_text_ids(prompt_embeds)
        return prompt_embeds, text_ids

    def _encode_vae_image(self, image: ms.Tensor, generator: np.random.Generator):
        if image.ndim != 4:
            raise ValueError(f"Expected image dims 4, got {image.ndim}.")

        image_latents = retrieve_latents(self.vae.encode(image), generator=generator, sample_mode="argmax")
        image_latents = self._patchify_latents(image_latents)

        latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(image_latents.dtype)
        latents_bn_std = mint.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps)
        image_latents = (image_latents - latents_bn_mean) / latents_bn_std

        return image_latents

    def prepare_latents(
        self,
        batch_size,
        num_latents_channels,
        height,
        width,
        dtype,
        generator: np.random.Generator,
        latents: Optional[ms.Tensor] = None,
    ):
        # VAE applies 8x compression on images but we must also account for packing which requires
        # latent height and width to be divisible by 2.
        height = 2 * (int(height) // (self.vae_scale_factor * 2))
        width = 2 * (int(width) // (self.vae_scale_factor * 2))

        shape = (batch_size, num_latents_channels * 4, height // 2, width // 2)
        if isinstance(generator, list) and len(generator) != batch_size:
            raise ValueError(
                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
            )
        if latents is None:
            latents = randn_tensor(shape, generator=generator, dtype=dtype)
        else:
            latents = latents.to(dtype=dtype)

        latent_ids = self._prepare_latent_ids(latents)

        latents = self._pack_latents(latents)  # [B, C, H, W] -> [B, H*W, C]
        return latents, latent_ids

    def prepare_image_latents(
        self,
        images: List[ms.Tensor],
        batch_size,
        generator: np.random.Generator,
        dtype,
    ):
        image_latents = []
        for image in images:
            image = image.to(dtype=dtype)
            imagge_latent = self._encode_vae_image(image=image, generator=generator)
            image_latents.append(imagge_latent)  # (1, 128, 32, 32)

        image_latent_ids = self._prepare_image_ids(image_latents)

        # Pack each latent and concatenate
        packed_latents = []
        for latent in image_latents:
            # latent: (1, 128, 32, 32)
            packed = self._pack_latents(latent)  # (1, 1024, 128)
            packed = packed.squeeze(0)  # (1024, 128) - remove batch dim
            packed_latents.append(packed)

        # Concatenate all reference tokens along sequence dimension
        image_latents = mint.cat(packed_latents, dim=0)  # (N*1024, 128)
        image_latents = image_latents.unsqueeze(0)  # (1, N*1024, 128)

        image_latents = image_latents.repeat(batch_size, 1, 1)
        image_latent_ids = image_latent_ids.repeat(batch_size, 1, 1)

        return image_latents, image_latent_ids

    def check_inputs(
        self,
        prompt,
        height,
        width,
        prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
    ):
        if (
            height is not None
            and height % (self.vae_scale_factor * 2) != 0
            or width is not None
            and width % (self.vae_scale_factor * 2) != 0
        ):
            logger.warning(
                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
            )

        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
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")

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

    @property
    def joint_attention_kwargs(self):
        return self._joint_attention_kwargs

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

    @property
    def current_timestep(self):
        return self._current_timestep

    @property
    def interrupt(self):
        return self._interrupt

    @ms._no_grad()
    def __call__(
        self,
        image: Optional[Union[List[PIL.Image.Image], PIL.Image.Image]] = None,
        prompt: Union[str, List[str]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        sigmas: Optional[List[float]] = None,
        guidance_scale: Optional[float] = 4.0,
        num_images_per_prompt: int = 1,
        generator: Optional[Union[np.random.Generator, List[np.random.Generator]]] = None,
        latents: Optional[ms.Tensor] = None,
        prompt_embeds: Optional[ms.Tensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        attention_kwargs: Optional[Dict[str, Any]] = None,
        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
        max_sequence_length: int = 512,
        text_encoder_out_layers: Tuple[int] = (10, 20, 30),
        caption_upsample_temperature: float = None,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        Args:
            image (`ms.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[ms.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
                `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
                numpy array and mindspore tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
                or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
                list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
                latents as `image`, but if passing latents directly it is not encoded again.
            prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
                instead.
            guidance_scale (`float`, *optional*, defaults to 1.0):
                Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
                a model to generate images more aligned with `prompt` at the expense of lower image quality.

                Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
                the [paper](https://huggingface.co/papers/2210.03142) to learn more.
            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The height in pixels of the generated image. This is set to 1024 by default for the best results.
            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
                The width in pixels of the generated image. This is set to 1024 by default for the best results.
            num_inference_steps (`int`, *optional*, defaults to 50):
                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
                expense of slower inference.
            sigmas (`List[float]`, *optional*):
                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            generator (`np.random.Generator` or `List[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.
            latents (`ms.Tensor`, *optional*):
                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
                tensor will be generated by sampling using the supplied random `generator`.
            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.
            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 `True`):
                Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
            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).
            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.
            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
            text_encoder_out_layers (`Tuple[int]`):
                Layer indices to use in the `text_encoder` to derive the final prompt embeddings.
            caption_upsample_temperature (`float`):
                When specified, we will try to perform caption upsampling for potentially improved outputs. We
                recommend setting it to 0.15 if caption upsampling is to be performed.

        Examples:

        Returns:
            [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`: [`~pipelines.flux2.Flux2PipelineOutput`] if
            `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
            generated images.
        """

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt=prompt,
            height=height,
            width=width,
            prompt_embeds=prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
        )

        self._guidance_scale = guidance_scale
        self._attention_kwargs = attention_kwargs
        self._current_timestep = None
        self._interrupt = False

        # 2. Define call parameters
        if prompt is not None and isinstance(prompt, str):
            batch_size = 1
        elif prompt is not None and isinstance(prompt, list):
            batch_size = len(prompt)
        else:
            batch_size = prompt_embeds.shape[0]

        # 3. prepare text embeddings
        if caption_upsample_temperature:
            prompt = self.upsample_prompt(prompt, images=image, temperature=caption_upsample_temperature)
        prompt_embeds, text_ids = self.encode_prompt(
            prompt=prompt,
            prompt_embeds=prompt_embeds,
            num_images_per_prompt=num_images_per_prompt,
            max_sequence_length=max_sequence_length,
            text_encoder_out_layers=text_encoder_out_layers,
        )

        # 4. process images
        if image is not None and not isinstance(image, list):
            image = [image]

        condition_images = None
        if image is not None:
            for img in image:
                self.image_processor.check_image_input(img)

            condition_images = []
            for img in image:
                image_width, image_height = img.size
                if image_width * image_height > 1024 * 1024:
                    img = self.image_processor._resize_to_target_area(img, 1024 * 1024)
                    image_width, image_height = img.size

                multiple_of = self.vae_scale_factor * 2
                image_width = (image_width // multiple_of) * multiple_of
                image_height = (image_height // multiple_of) * multiple_of
                img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop")
                condition_images.append(img)
                height = height or image_height
                width = width or image_width

        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

        # 5. prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_ids = self.prepare_latents(
            batch_size=batch_size * num_images_per_prompt,
            num_latents_channels=num_channels_latents,
            height=height,
            width=width,
            dtype=prompt_embeds.dtype,
            generator=generator,
            latents=latents,
        )

        image_latents = None
        image_latent_ids = None
        if condition_images is not None:
            image_latents, image_latent_ids = self.prepare_image_latents(
                images=condition_images,
                batch_size=batch_size * num_images_per_prompt,
                generator=generator,
                dtype=self.vae.dtype,
            )

        # 6. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
        if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas:
            sigmas = None
        image_seq_len = latents.shape[1]
        mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps)
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            sigmas=sigmas,
            mu=mu,
        )
        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

        # handle guidance
        guidance = ms.ops.full([1], guidance_scale, dtype=ms.float32)
        guidance = guidance.expand((latents.shape[0],))

        # 7. Denoising loop
        # We set the index here to remove DtoH sync, helpful especially during compilation.
        # Check out more details here: https://github.com/huggingface/diffusers/pull/11696
        self.scheduler.set_begin_index(0)
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                self._current_timestep = t
                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand((latents.shape[0],)).to(latents.dtype)

                latent_model_input = latents.to(self.transformer.dtype)
                latent_image_ids = latent_ids

                if image_latents is not None:
                    latent_model_input = mint.cat([latents, image_latents], dim=1).to(self.transformer.dtype)
                    latent_image_ids = mint.cat([latent_ids, image_latent_ids], dim=1)

                noise_pred = self.transformer(
                    hidden_states=latent_model_input,  # (B, image_seq_len, C)
                    timestep=timestep / 1000,
                    guidance=guidance,
                    encoder_hidden_states=prompt_embeds,
                    txt_ids=text_ids,  # B, text_seq_len, 4
                    img_ids=latent_image_ids,  # B, image_seq_len, 4
                    joint_attention_kwargs=self._attention_kwargs,
                    return_dict=False,
                )[0]

                noise_pred = noise_pred[:, : latents.shape[1] :]

                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]

                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)

                # 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()

        self._current_timestep = None

        if output_type == "latent":
            image = latents
        else:
            latents = self._unpack_latents_with_ids(latents, latent_ids)

            latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.dtype)
            latents_bn_std = mint.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(
                latents.dtype
            )
            latents = latents * latents_bn_std + latents_bn_mean
            latents = self._unpatchify_latents(latents)

            image = self.vae.decode(latents, return_dict=False)[0]
            image = self.image_processor.postprocess(image, output_type=output_type)

        if not return_dict:
            return (image,)

        return Flux2PipelineOutput(images=image)

mindone.diffusers.Flux2Pipeline.__call__(image=None, prompt=None, height=None, width=None, num_inference_steps=50, sigmas=None, guidance_scale=4.0, num_images_per_prompt=1, generator=None, latents=None, prompt_embeds=None, output_type='pil', return_dict=True, attention_kwargs=None, callback_on_step_end=None, callback_on_step_end_tensor_inputs=['latents'], max_sequence_length=512, text_encoder_out_layers=(10, 20, 30), caption_upsample_temperature=None)

Function invoked when calling the pipeline for generation.

PARAMETER	DESCRIPTION
image	Image, numpy array or tensor representing an image batch to be used as the starting point. For both numpy array and mindspore tensor, the expected value range is between [0, 1] If it's a tensor or a list or tensors, the expected shape should be (B, C, H, W) or (C, H, W). If it is a numpy array or a list of arrays, the expected shape should be (B, H, W, C) or (H, W, C) It can also accept image latents as image, but if passing latents directly it is not encoded again. TYPE: `ms.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[ms.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]` DEFAULT: None
prompt	The prompt or prompts to guide the image generation. If not defined, one has to pass prompt_embeds. instead. TYPE: `str` or `List[str]`, *optional* DEFAULT: None
guidance_scale	Embedded guiddance scale is enabled by setting guidance_scale > 1. Higher guidance_scale encourages a model to generate images more aligned with prompt at the expense of lower image quality. Guidance-distilled models approximates true classifer-free guidance for guidance_scale > 1. Refer to the paper to learn more. TYPE: `float`, *optional*, defaults to 1.0 DEFAULT: 4.0
height	The height in pixels of the generated image. This is set to 1024 by default for the best results. TYPE: `int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor DEFAULT: None
width	The width in pixels of the generated image. This is set to 1024 by default for the best results. TYPE: `int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor DEFAULT: None
num_inference_steps	The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference. TYPE: `int`, *optional*, defaults to 50 DEFAULT: 50
sigmas	Custom sigmas to use for the denoising process with schedulers which support a sigmas argument in their set_timesteps method. If not defined, the default behavior when num_inference_steps is passed will be used. TYPE: `List[float]`, *optional* DEFAULT: None
num_images_per_prompt	The number of images to generate per prompt. TYPE: `int`, *optional*, defaults to 1 DEFAULT: 1
generator	One or a list of np.random.Generator(s) to make generation deterministic. TYPE: `np.random.Generator` or `List[np.random.Generator]`, *optional* DEFAULT: None
latents	Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image generation. Can be used to tweak the same generation with different prompts. If not provided, a latents tensor will be generated by sampling using the supplied random generator. TYPE: `ms.Tensor`, *optional* DEFAULT: None
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
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.qwenimage.QwenImagePipelineOutput] instead of a plain tuple. TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True
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
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']
max_sequence_length	Maximum sequence length to use with the prompt. TYPE: `int` defaults to 512 DEFAULT: 512
text_encoder_out_layers	Layer indices to use in the text_encoder to derive the final prompt embeddings. TYPE: `Tuple[int]` DEFAULT: (10, 20, 30)
caption_upsample_temperature	When specified, we will try to perform caption upsampling for potentially improved outputs. We recommend setting it to 0.15 if caption upsampling is to be performed. TYPE: `float` DEFAULT: None

RETURNS	DESCRIPTION
	[~pipelines.flux2.Flux2PipelineOutput] or tuple: [~pipelines.flux2.Flux2PipelineOutput] 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/flux2/pipeline_flux2.py

@ms._no_grad()
def __call__(
    self,
    image: Optional[Union[List[PIL.Image.Image], PIL.Image.Image]] = None,
    prompt: Union[str, List[str]] = None,
    height: Optional[int] = None,
    width: Optional[int] = None,
    num_inference_steps: int = 50,
    sigmas: Optional[List[float]] = None,
    guidance_scale: Optional[float] = 4.0,
    num_images_per_prompt: int = 1,
    generator: Optional[Union[np.random.Generator, List[np.random.Generator]]] = None,
    latents: Optional[ms.Tensor] = None,
    prompt_embeds: Optional[ms.Tensor] = None,
    output_type: Optional[str] = "pil",
    return_dict: bool = True,
    attention_kwargs: Optional[Dict[str, Any]] = None,
    callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
    callback_on_step_end_tensor_inputs: List[str] = ["latents"],
    max_sequence_length: int = 512,
    text_encoder_out_layers: Tuple[int] = (10, 20, 30),
    caption_upsample_temperature: float = None,
):
    r"""
    Function invoked when calling the pipeline for generation.

    Args:
        image (`ms.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[ms.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
            `Image`, numpy array or tensor representing an image batch to be used as the starting point. For both
            numpy array and mindspore tensor, the expected value range is between `[0, 1]` If it's a tensor or a list
            or tensors, the expected shape should be `(B, C, H, W)` or `(C, H, W)`. If it is a numpy array or a
            list of arrays, the expected shape should be `(B, H, W, C)` or `(H, W, C)` It can also accept image
            latents as `image`, but if passing latents directly it is not encoded again.
        prompt (`str` or `List[str]`, *optional*):
            The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
            instead.
        guidance_scale (`float`, *optional*, defaults to 1.0):
            Embedded guiddance scale is enabled by setting `guidance_scale` > 1. Higher `guidance_scale` encourages
            a model to generate images more aligned with `prompt` at the expense of lower image quality.

            Guidance-distilled models approximates true classifer-free guidance for `guidance_scale` > 1. Refer to
            the [paper](https://huggingface.co/papers/2210.03142) to learn more.
        height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
            The height in pixels of the generated image. This is set to 1024 by default for the best results.
        width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
            The width in pixels of the generated image. This is set to 1024 by default for the best results.
        num_inference_steps (`int`, *optional*, defaults to 50):
            The number of denoising steps. More denoising steps usually lead to a higher quality image at the
            expense of slower inference.
        sigmas (`List[float]`, *optional*):
            Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
            their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
            will be used.
        num_images_per_prompt (`int`, *optional*, defaults to 1):
            The number of images to generate per prompt.
        generator (`np.random.Generator` or `List[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.
        latents (`ms.Tensor`, *optional*):
            Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
            generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
            tensor will be generated by sampling using the supplied random `generator`.
        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.
        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 `True`):
            Whether or not to return a [`~pipelines.qwenimage.QwenImagePipelineOutput`] instead of a plain tuple.
        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).
        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.
        max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
        text_encoder_out_layers (`Tuple[int]`):
            Layer indices to use in the `text_encoder` to derive the final prompt embeddings.
        caption_upsample_temperature (`float`):
            When specified, we will try to perform caption upsampling for potentially improved outputs. We
            recommend setting it to 0.15 if caption upsampling is to be performed.

    Examples:

    Returns:
        [`~pipelines.flux2.Flux2PipelineOutput`] or `tuple`: [`~pipelines.flux2.Flux2PipelineOutput`] if
        `return_dict` is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the
        generated images.
    """

    # 1. Check inputs. Raise error if not correct
    self.check_inputs(
        prompt=prompt,
        height=height,
        width=width,
        prompt_embeds=prompt_embeds,
        callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
    )

    self._guidance_scale = guidance_scale
    self._attention_kwargs = attention_kwargs
    self._current_timestep = None
    self._interrupt = False

    # 2. Define call parameters
    if prompt is not None and isinstance(prompt, str):
        batch_size = 1
    elif prompt is not None and isinstance(prompt, list):
        batch_size = len(prompt)
    else:
        batch_size = prompt_embeds.shape[0]

    # 3. prepare text embeddings
    if caption_upsample_temperature:
        prompt = self.upsample_prompt(prompt, images=image, temperature=caption_upsample_temperature)
    prompt_embeds, text_ids = self.encode_prompt(
        prompt=prompt,
        prompt_embeds=prompt_embeds,
        num_images_per_prompt=num_images_per_prompt,
        max_sequence_length=max_sequence_length,
        text_encoder_out_layers=text_encoder_out_layers,
    )

    # 4. process images
    if image is not None and not isinstance(image, list):
        image = [image]

    condition_images = None
    if image is not None:
        for img in image:
            self.image_processor.check_image_input(img)

        condition_images = []
        for img in image:
            image_width, image_height = img.size
            if image_width * image_height > 1024 * 1024:
                img = self.image_processor._resize_to_target_area(img, 1024 * 1024)
                image_width, image_height = img.size

            multiple_of = self.vae_scale_factor * 2
            image_width = (image_width // multiple_of) * multiple_of
            image_height = (image_height // multiple_of) * multiple_of
            img = self.image_processor.preprocess(img, height=image_height, width=image_width, resize_mode="crop")
            condition_images.append(img)
            height = height or image_height
            width = width or image_width

    height = height or self.default_sample_size * self.vae_scale_factor
    width = width or self.default_sample_size * self.vae_scale_factor

    # 5. prepare latent variables
    num_channels_latents = self.transformer.config.in_channels // 4
    latents, latent_ids = self.prepare_latents(
        batch_size=batch_size * num_images_per_prompt,
        num_latents_channels=num_channels_latents,
        height=height,
        width=width,
        dtype=prompt_embeds.dtype,
        generator=generator,
        latents=latents,
    )

    image_latents = None
    image_latent_ids = None
    if condition_images is not None:
        image_latents, image_latent_ids = self.prepare_image_latents(
            images=condition_images,
            batch_size=batch_size * num_images_per_prompt,
            generator=generator,
            dtype=self.vae.dtype,
        )

    # 6. Prepare timesteps
    sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
    if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas:
        sigmas = None
    image_seq_len = latents.shape[1]
    mu = compute_empirical_mu(image_seq_len=image_seq_len, num_steps=num_inference_steps)
    timesteps, num_inference_steps = retrieve_timesteps(
        self.scheduler,
        num_inference_steps,
        sigmas=sigmas,
        mu=mu,
    )
    num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
    self._num_timesteps = len(timesteps)

    # handle guidance
    guidance = ms.ops.full([1], guidance_scale, dtype=ms.float32)
    guidance = guidance.expand((latents.shape[0],))

    # 7. Denoising loop
    # We set the index here to remove DtoH sync, helpful especially during compilation.
    # Check out more details here: https://github.com/huggingface/diffusers/pull/11696
    self.scheduler.set_begin_index(0)
    with self.progress_bar(total=num_inference_steps) as progress_bar:
        for i, t in enumerate(timesteps):
            if self.interrupt:
                continue

            self._current_timestep = t
            # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
            timestep = t.expand((latents.shape[0],)).to(latents.dtype)

            latent_model_input = latents.to(self.transformer.dtype)
            latent_image_ids = latent_ids

            if image_latents is not None:
                latent_model_input = mint.cat([latents, image_latents], dim=1).to(self.transformer.dtype)
                latent_image_ids = mint.cat([latent_ids, image_latent_ids], dim=1)

            noise_pred = self.transformer(
                hidden_states=latent_model_input,  # (B, image_seq_len, C)
                timestep=timestep / 1000,
                guidance=guidance,
                encoder_hidden_states=prompt_embeds,
                txt_ids=text_ids,  # B, text_seq_len, 4
                img_ids=latent_image_ids,  # B, image_seq_len, 4
                joint_attention_kwargs=self._attention_kwargs,
                return_dict=False,
            )[0]

            noise_pred = noise_pred[:, : latents.shape[1] :]

            # compute the previous noisy sample x_t -> x_t-1
            latents_dtype = latents.dtype
            latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]

            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)

            # 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()

    self._current_timestep = None

    if output_type == "latent":
        image = latents
    else:
        latents = self._unpack_latents_with_ids(latents, latent_ids)

        latents_bn_mean = self.vae.bn.running_mean.view(1, -1, 1, 1).to(latents.dtype)
        latents_bn_std = mint.sqrt(self.vae.bn.running_var.view(1, -1, 1, 1) + self.vae.config.batch_norm_eps).to(
            latents.dtype
        )
        latents = latents * latents_bn_std + latents_bn_mean
        latents = self._unpatchify_latents(latents)

        image = self.vae.decode(latents, return_dict=False)[0]
        image = self.image_processor.postprocess(image, output_type=output_type)

    if not return_dict:
        return (image,)

    return Flux2PipelineOutput(images=image)