UNet2DConditionModel

The UNet model was originally introduced by Ronneberger et al. for biomedical image segmentation, but it is also commonly used in 🤗 Diffusers because it outputs images that are the same size as the input. It is one of the most important components of a diffusion system because it facilitates the actual diffusion process. There are several variants of the UNet model in 🤗 Diffusers, depending on it's number of dimensions and whether it is a conditional model or not. This is a 2D UNet conditional model.

The abstract from the paper is:

There is large consent that successful training of deep networks requires many thousand annotated training samples. In this paper, we present a network and training strategy that relies on the strong use of data augmentation to use the available annotated samples more efficiently. The architecture consists of a contracting path to capture context and a symmetric expanding path that enables precise localization. We show that such a network can be trained end-to-end from very few images and outperforms the prior best method (a sliding-window convolutional network) on the ISBI challenge for segmentation of neuronal structures in electron microscopic stacks. Using the same network trained on transmitted light microscopy images (phase contrast and DIC) we won the ISBI cell tracking challenge 2015 in these categories by a large margin. Moreover, the network is fast. Segmentation of a 512x512 image takes less than a second on a recent GPU. The full implementation (based on Caffe) and the trained networks are available at http://lmb.informatik.uni-freiburg.de/people/ronneber/u-net.

mindone.diffusers.UNet2DConditionModel

Bases: ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin

A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample shaped output.

This model inherits from [ModelMixin]. Check the superclass documentation for it's generic methods implemented for all models (such as downloading or saving).

PARAMETER	DESCRIPTION
sample_size	Height and width of input/output sample. TYPE: `int` or `Tuple[int, int]`, *optional*, defaults to `None` DEFAULT: None
in_channels	Number of channels in the input sample. TYPE: `int`, *optional*, defaults to 4 DEFAULT: 4
out_channels	Number of channels in the output. TYPE: `int`, *optional*, defaults to 4 DEFAULT: 4
center_input_sample	Whether to center the input sample. TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False
flip_sin_to_cos	Whether to flip the sin to cos in the time embedding. TYPE: `bool`, *optional*, defaults to `True` DEFAULT: True
freq_shift	The frequency shift to apply to the time embedding. TYPE: `int`, *optional*, defaults to 0 DEFAULT: 0
down_block_types	The tuple of downsample blocks to use. TYPE: `Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")` DEFAULT: ('CrossAttnDownBlock2D', 'CrossAttnDownBlock2D', 'CrossAttnDownBlock2D', 'DownBlock2D')
mid_block_type	Block type for middle of UNet, it can be one of UNetMidBlock2DCrossAttn, UNetMidBlock2D, or UNetMidBlock2DSimpleCrossAttn. If None, the mid block layer is skipped. TYPE: `str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"` DEFAULT: 'UNetMidBlock2DCrossAttn'
up_block_types	The tuple of upsample blocks to use. TYPE: `Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")` DEFAULT: ('UpBlock2D', 'CrossAttnUpBlock2D', 'CrossAttnUpBlock2D', 'CrossAttnUpBlock2D')
only_cross_attention(`bool`	Whether to include self-attention in the basic transformer blocks, see [~models.attention.BasicTransformerBlock]. TYPE: or `Tuple[bool]`, *optional*, default to `False`
block_out_channels	The tuple of output channels for each block. TYPE: `Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)` DEFAULT: (320, 640, 1280, 1280)
layers_per_block	The number of layers per block. TYPE: `int`, *optional*, defaults to 2 DEFAULT: 2
downsample_padding	The padding to use for the downsampling convolution. TYPE: `int`, *optional*, defaults to 1 DEFAULT: 1
mid_block_scale_factor	The scale factor to use for the mid block. TYPE: `float`, *optional*, defaults to 1.0 DEFAULT: 1
dropout	The dropout probability to use. TYPE: `float`, *optional*, defaults to 0.0 DEFAULT: 0.0
act_fn	The activation function to use. TYPE: `str`, *optional*, defaults to `"silu"` DEFAULT: 'silu'
norm_num_groups	The number of groups to use for the normalization. If None, normalization and activation layers is skipped in post-processing. TYPE: `int`, *optional*, defaults to 32 DEFAULT: 32
norm_eps	The epsilon to use for the normalization. TYPE: `float`, *optional*, defaults to 1e-5 DEFAULT: 1e-05
cross_attention_dim	The dimension of the cross attention features. TYPE: `int` or `Tuple[int]`, *optional*, defaults to 1280 DEFAULT: 1280
transformer_layers_per_block	The number of transformer blocks of type [~models.attention.BasicTransformerBlock]. Only relevant for [~models.unets.unet_2d_blocks.CrossAttnDownBlock2D], [~models.unets.unet_2d_blocks.CrossAttnUpBlock2D], [~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn]. TYPE: `int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1 DEFAULT: 1
reverse_transformer_layers_per_block	(Tuple[Tuple], optional, defaults to None): The number of transformer blocks of type [~models.attention.BasicTransformerBlock], in the upsampling blocks of the U-Net. Only relevant if transformer_layers_per_block is of type Tuple[Tuple] and for [~models.unets.unet_2d_blocks.CrossAttnDownBlock2D], [~models.unets.unet_2d_blocks.CrossAttnUpBlock2D], [~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn]. DEFAULT: None
encoder_hid_dim	If encoder_hid_dim_type is defined, encoder_hidden_states will be projected from encoder_hid_dim dimension to cross_attention_dim. TYPE: `int`, *optional*, defaults to None DEFAULT: None
encoder_hid_dim_type	If given, the encoder_hidden_states and potentially other embeddings are down-projected to text embeddings of dimension cross_attention according to encoder_hid_dim_type. TYPE: `str`, *optional*, defaults to `None` DEFAULT: None
attention_head_dim	The dimension of the attention heads. TYPE: `int`, *optional*, defaults to 8 DEFAULT: 8
num_attention_heads	The number of attention heads. If not defined, defaults to attention_head_dim TYPE: `int`, *optional* DEFAULT: None
resnet_time_scale_shift	Time scale shift config for ResNet blocks (see [~models.resnet.ResnetBlock2D]). Choose from default or scale_shift. TYPE: `str`, *optional*, defaults to `"default"` DEFAULT: 'default'
class_embed_type	The type of class embedding to use which is ultimately summed with the time embeddings. Choose from None, "timestep", "identity", "projection", or "simple_projection". TYPE: `str`, *optional*, defaults to `None` DEFAULT: None
addition_embed_type	Configures an optional embedding which will be summed with the time embeddings. Choose from None or "text". "text" will use the TextTimeEmbedding layer. TYPE: `str`, *optional*, defaults to `None` DEFAULT: None
addition_time_embed_dim	(int, optional, defaults to None): Dimension for the timestep embeddings. TYPE: Optional[int] DEFAULT: None
num_class_embeds	Input dimension of the learnable embedding matrix to be projected to time_embed_dim, when performing class conditioning with class_embed_type equal to None. TYPE: `int`, *optional*, defaults to `None` DEFAULT: None
time_embedding_type	The type of position embedding to use for timesteps. Choose from positional or fourier. TYPE: `str`, *optional*, defaults to `positional` DEFAULT: 'positional'
time_embedding_dim	An optional override for the dimension of the projected time embedding. TYPE: `int`, *optional*, defaults to `None` DEFAULT: None
time_embedding_act_fn	Optional activation function to use only once on the time embeddings before they are passed to the rest of the UNet. Choose from silu, mish, gelu, and swish. TYPE: `str`, *optional*, defaults to `None` DEFAULT: None
timestep_post_act	The second activation function to use in timestep embedding. Choose from silu, mish and gelu. TYPE: `str`, *optional*, defaults to `None` DEFAULT: None
time_cond_proj_dim	The dimension of cond_proj layer in the timestep embedding. TYPE: `int`, *optional*, defaults to `None` DEFAULT: None
conv_in_kernel	The kernel size of conv_in layer. TYPE: `int`, *optional*, default to `3` DEFAULT: 3
conv_out_kernel	The kernel size of conv_out layer. TYPE: `int`, *optional*, default to `3` DEFAULT: 3
projection_class_embeddings_input_dim	The dimension of the class_labels input when class_embed_type="projection". Required when class_embed_type="projection". TYPE: `int`, *optional* DEFAULT: None
class_embeddings_concat	Whether to concatenate the time embeddings with the class embeddings. TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False
mid_block_only_cross_attention	Whether to use cross attention with the mid block when using the UNetMidBlock2DSimpleCrossAttn. If only_cross_attention is given as a single boolean and mid_block_only_cross_attention is None, the only_cross_attention value is used as the value for mid_block_only_cross_attention. Default to False otherwise. TYPE: `bool`, *optional*, defaults to `None` DEFAULT: None

Source code in mindone/diffusers/models/unets/unet_2d_condition.py

class UNet2DConditionModel(
    ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
):
    r"""
    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
    shaped output.

    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
    for all models (such as downloading or saving).

    Parameters:
        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
            Height and width of input/output sample.
        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
            Whether to flip the sin to cos in the time embedding.
        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
            The tuple of downsample blocks to use.
        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
            The tuple of upsample blocks to use.
        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
            Whether to include self-attention in the basic transformer blocks, see
            [`~models.attention.BasicTransformerBlock`].
        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
            The tuple of output channels for each block.
        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
            If `None`, normalization and activation layers is skipped in post-processing.
        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
            The dimension of the cross attention features.
        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
            [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
            [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
        reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
            [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
            [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
        encoder_hid_dim (`int`, *optional*, defaults to None):
            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
            dimension to `cross_attention_dim`.
        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
        num_attention_heads (`int`, *optional*):
            The number of attention heads. If not defined, defaults to `attention_head_dim`
        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
        class_embed_type (`str`, *optional*, defaults to `None`):
            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
        addition_embed_type (`str`, *optional*, defaults to `None`):
            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
            "text". "text" will use the `TextTimeEmbedding` layer.
        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
            Dimension for the timestep embeddings.
        num_class_embeds (`int`, *optional*, defaults to `None`):
            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
            class conditioning with `class_embed_type` equal to `None`.
        time_embedding_type (`str`, *optional*, defaults to `positional`):
            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
        time_embedding_dim (`int`, *optional*, defaults to `None`):
            An optional override for the dimension of the projected time embedding.
        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
            Optional activation function to use only once on the time embeddings before they are passed to the rest of
            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
        timestep_post_act (`str`, *optional*, defaults to `None`):
            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
            The dimension of `cond_proj` layer in the timestep embedding.
        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
            embeddings with the class embeddings.
        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
            otherwise.
    """

    _supports_gradient_checkpointing = True
    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]

    @register_to_config
    def __init__(
        self,
        sample_size: Optional[int] = None,
        in_channels: int = 4,
        out_channels: int = 4,
        center_input_sample: bool = False,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        down_block_types: Tuple[str] = (
            "CrossAttnDownBlock2D",
            "CrossAttnDownBlock2D",
            "CrossAttnDownBlock2D",
            "DownBlock2D",
        ),
        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
        only_cross_attention: Union[bool, Tuple[bool]] = False,
        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
        layers_per_block: Union[int, Tuple[int]] = 2,
        downsample_padding: int = 1,
        mid_block_scale_factor: float = 1,
        dropout: float = 0.0,
        act_fn: str = "silu",
        norm_num_groups: Optional[int] = 32,
        norm_eps: float = 1e-5,
        cross_attention_dim: Union[int, Tuple[int]] = 1280,
        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
        encoder_hid_dim: Optional[int] = None,
        encoder_hid_dim_type: Optional[str] = None,
        attention_head_dim: Union[int, Tuple[int]] = 8,
        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
        dual_cross_attention: bool = False,
        use_linear_projection: bool = False,
        class_embed_type: Optional[str] = None,
        addition_embed_type: Optional[str] = None,
        addition_time_embed_dim: Optional[int] = None,
        num_class_embeds: Optional[int] = None,
        upcast_attention: bool = False,
        resnet_time_scale_shift: str = "default",
        resnet_skip_time_act: bool = False,
        resnet_out_scale_factor: float = 1.0,
        time_embedding_type: str = "positional",
        time_embedding_dim: Optional[int] = None,
        time_embedding_act_fn: Optional[str] = None,
        timestep_post_act: Optional[str] = None,
        time_cond_proj_dim: Optional[int] = None,
        conv_in_kernel: int = 3,
        conv_out_kernel: int = 3,
        projection_class_embeddings_input_dim: Optional[int] = None,
        attention_type: str = "default",
        class_embeddings_concat: bool = False,
        mid_block_only_cross_attention: Optional[bool] = None,
        cross_attention_norm: Optional[str] = None,
        addition_embed_type_num_heads: int = 64,
    ):
        super().__init__()

        self.sample_size = sample_size

        if num_attention_heads is not None:
            raise ValueError(
                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."  # noqa: E501
            )

        # If `num_attention_heads` is not defined (which is the case for most models)
        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
        # The reason for this behavior is to correct for incorrectly named variables that were introduced
        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131  # noqa: E501
        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
        # which is why we correct for the naming here.
        num_attention_heads = num_attention_heads or attention_head_dim

        # Check inputs
        self._check_config(
            down_block_types=down_block_types,
            up_block_types=up_block_types,
            only_cross_attention=only_cross_attention,
            block_out_channels=block_out_channels,
            layers_per_block=layers_per_block,
            cross_attention_dim=cross_attention_dim,
            transformer_layers_per_block=transformer_layers_per_block,
            reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
            attention_head_dim=attention_head_dim,
            num_attention_heads=num_attention_heads,
        )

        # input
        conv_in_padding = (conv_in_kernel - 1) // 2
        self.conv_in = nn.Conv2d(
            in_channels,
            block_out_channels[0],
            kernel_size=conv_in_kernel,
            pad_mode="pad",
            padding=conv_in_padding,
            has_bias=True,
        )

        # time
        time_embed_dim, timestep_input_dim = self._set_time_proj(
            time_embedding_type,
            block_out_channels=block_out_channels,
            flip_sin_to_cos=flip_sin_to_cos,
            freq_shift=freq_shift,
            time_embedding_dim=time_embedding_dim,
        )

        self.time_embedding = TimestepEmbedding(
            timestep_input_dim,
            time_embed_dim,
            act_fn=act_fn,
            post_act_fn=timestep_post_act,
            cond_proj_dim=time_cond_proj_dim,
        )

        self._set_encoder_hid_proj(
            encoder_hid_dim_type,
            cross_attention_dim=cross_attention_dim,
            encoder_hid_dim=encoder_hid_dim,
        )

        self.has_text_encoder_hid_proj = (
            False  # Used in `process_encoder_hidden_states`, Kolors Unet already has a `encoder_hid_proj`
        )

        # class embedding
        self._set_class_embedding(
            class_embed_type,
            act_fn=act_fn,
            num_class_embeds=num_class_embeds,
            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
            time_embed_dim=time_embed_dim,
            timestep_input_dim=timestep_input_dim,
        )

        self._set_add_embedding(
            addition_embed_type,
            addition_embed_type_num_heads=addition_embed_type_num_heads,
            addition_time_embed_dim=addition_time_embed_dim,
            cross_attention_dim=cross_attention_dim,
            encoder_hid_dim=encoder_hid_dim,
            flip_sin_to_cos=flip_sin_to_cos,
            freq_shift=freq_shift,
            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
            time_embed_dim=time_embed_dim,
        )

        if time_embedding_act_fn is None:
            self.time_embed_act = None
        else:
            self.time_embed_act = get_activation(time_embedding_act_fn)()

        if isinstance(only_cross_attention, bool):
            if mid_block_only_cross_attention is None:
                mid_block_only_cross_attention = only_cross_attention

            only_cross_attention = [only_cross_attention] * len(down_block_types)

        if mid_block_only_cross_attention is None:
            mid_block_only_cross_attention = False

        if isinstance(num_attention_heads, int):
            num_attention_heads = (num_attention_heads,) * len(down_block_types)

        if isinstance(attention_head_dim, int):
            attention_head_dim = (attention_head_dim,) * len(down_block_types)

        if isinstance(cross_attention_dim, int):
            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)

        if isinstance(layers_per_block, int):
            layers_per_block = [layers_per_block] * len(down_block_types)

        if isinstance(transformer_layers_per_block, int):
            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)

        if class_embeddings_concat:
            # The time embeddings are concatenated with the class embeddings. The dimension of the
            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
            # regular time embeddings
            blocks_time_embed_dim = time_embed_dim * 2
        else:
            blocks_time_embed_dim = time_embed_dim

        self.down_blocks = nn.CellList()
        self.up_blocks = nn.CellList()
        # down
        down_blocks = []
        output_channel = block_out_channels[0]
        for i, down_block_type in enumerate(down_block_types):
            input_channel = output_channel
            output_channel = block_out_channels[i]
            is_final_block = i == len(block_out_channels) - 1

            down_block = get_down_block(
                down_block_type,
                num_layers=layers_per_block[i],
                transformer_layers_per_block=transformer_layers_per_block[i],
                in_channels=input_channel,
                out_channels=output_channel,
                temb_channels=blocks_time_embed_dim,
                add_downsample=not is_final_block,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resnet_groups=norm_num_groups,
                cross_attention_dim=cross_attention_dim[i],
                num_attention_heads=num_attention_heads[i],
                downsample_padding=downsample_padding,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                only_cross_attention=only_cross_attention[i],
                upcast_attention=upcast_attention,
                resnet_time_scale_shift=resnet_time_scale_shift,
                attention_type=attention_type,
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
                cross_attention_norm=cross_attention_norm,
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
                dropout=dropout,
            )
            down_blocks.append(down_block)
        self.down_blocks = nn.CellList(down_blocks)

        # mid
        self.mid_block = get_mid_block(
            mid_block_type,
            temb_channels=blocks_time_embed_dim,
            in_channels=block_out_channels[-1],
            resnet_eps=norm_eps,
            resnet_act_fn=act_fn,
            resnet_groups=norm_num_groups,
            output_scale_factor=mid_block_scale_factor,
            transformer_layers_per_block=transformer_layers_per_block[-1],
            num_attention_heads=num_attention_heads[-1],
            cross_attention_dim=cross_attention_dim[-1],
            dual_cross_attention=dual_cross_attention,
            use_linear_projection=use_linear_projection,
            mid_block_only_cross_attention=mid_block_only_cross_attention,
            upcast_attention=upcast_attention,
            resnet_time_scale_shift=resnet_time_scale_shift,
            attention_type=attention_type,
            resnet_skip_time_act=resnet_skip_time_act,
            cross_attention_norm=cross_attention_norm,
            attention_head_dim=attention_head_dim[-1],
            dropout=dropout,
        )

        # count how many layers upsample the images
        self.num_upsamplers = 0

        # up
        up_blocks = []
        layers_per_resnet_in_up_blocks = []
        reversed_block_out_channels = list(reversed(block_out_channels))
        reversed_num_attention_heads = list(reversed(num_attention_heads))
        reversed_layers_per_block = list(reversed(layers_per_block))
        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
        reversed_transformer_layers_per_block = (
            list(reversed(transformer_layers_per_block))
            if reverse_transformer_layers_per_block is None
            else reverse_transformer_layers_per_block
        )
        only_cross_attention = list(reversed(only_cross_attention))

        output_channel = reversed_block_out_channels[0]
        for i, up_block_type in enumerate(up_block_types):
            is_final_block = i == len(block_out_channels) - 1

            prev_output_channel = output_channel
            output_channel = reversed_block_out_channels[i]
            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]

            # add upsample block for all BUT final layer
            if not is_final_block:
                add_upsample = True
                self.num_upsamplers += 1
            else:
                add_upsample = False

            up_block = get_up_block(
                up_block_type,
                num_layers=reversed_layers_per_block[i] + 1,
                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
                in_channels=input_channel,
                out_channels=output_channel,
                prev_output_channel=prev_output_channel,
                temb_channels=blocks_time_embed_dim,
                add_upsample=add_upsample,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                resolution_idx=i,
                resnet_groups=norm_num_groups,
                cross_attention_dim=reversed_cross_attention_dim[i],
                num_attention_heads=reversed_num_attention_heads[i],
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                only_cross_attention=only_cross_attention[i],
                upcast_attention=upcast_attention,
                resnet_time_scale_shift=resnet_time_scale_shift,
                attention_type=attention_type,
                resnet_skip_time_act=resnet_skip_time_act,
                resnet_out_scale_factor=resnet_out_scale_factor,
                cross_attention_norm=cross_attention_norm,
                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
                dropout=dropout,
            )
            up_blocks.append(up_block)
            prev_output_channel = output_channel
            layers_per_resnet_in_up_blocks.append(len(up_block.resnets))
        self.up_blocks = nn.CellList(up_blocks)
        self.layers_per_resnet_in_up_blocks = layers_per_resnet_in_up_blocks

        # out
        if norm_num_groups is not None:
            self.conv_norm_out = GroupNorm(num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps)

            self.conv_act = get_activation(act_fn)()

        else:
            self.conv_norm_out = None
            self.conv_act = None

        conv_out_padding = (conv_out_kernel - 1) // 2
        self.conv_out = nn.Conv2d(
            block_out_channels[0],
            out_channels,
            kernel_size=conv_out_kernel,
            pad_mode="pad",
            padding=conv_out_padding,
            has_bias=True,
        )

        self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)

        self.center_input_sample = self.config.center_input_sample
        self.class_embed_type = self.config.class_embed_type
        self.class_embeddings_concat = self.config.class_embeddings_concat
        self.addition_embed_type = self.config.addition_embed_type
        self.encoder_hid_dim_type = self.config.encoder_hid_dim_type

    def _check_config(
        self,
        down_block_types: Tuple[str],
        up_block_types: Tuple[str],
        only_cross_attention: Union[bool, Tuple[bool]],
        block_out_channels: Tuple[int],
        layers_per_block: Union[int, Tuple[int]],
        cross_attention_dim: Union[int, Tuple[int]],
        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
        reverse_transformer_layers_per_block: bool,
        attention_head_dim: int,
        num_attention_heads: Optional[Union[int, Tuple[int]]],
    ):
        if len(down_block_types) != len(up_block_types):
            raise ValueError(
                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."  # noqa: E501
            )

        if len(block_out_channels) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."  # noqa: E501
            )

        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."  # noqa: E501
            )

        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."  # noqa: E501
            )

        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."  # noqa: E501
            )

        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."  # noqa: E501
            )

        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
            raise ValueError(
                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."  # noqa: E501
            )
        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
            for layer_number_per_block in transformer_layers_per_block:
                if isinstance(layer_number_per_block, list):
                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")

    def _set_time_proj(
        self,
        time_embedding_type: str,
        block_out_channels: int,
        flip_sin_to_cos: bool,
        freq_shift: float,
        time_embedding_dim: int,
    ) -> Tuple[int, int]:
        if time_embedding_type == "fourier":
            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
            if time_embed_dim % 2 != 0:
                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
            self.time_proj = GaussianFourierProjection(
                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
            )
            timestep_input_dim = time_embed_dim
        elif time_embedding_type == "positional":
            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4

            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
            timestep_input_dim = block_out_channels[0]
        else:
            raise ValueError(
                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
            )
        return time_embed_dim, timestep_input_dim

    def _set_encoder_hid_proj(
        self,
        encoder_hid_dim_type: Optional[str],
        cross_attention_dim: Union[int, Tuple[int]],
        encoder_hid_dim: Optional[int],
    ):
        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
            encoder_hid_dim_type = "text_proj"
            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")

        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
            raise ValueError(
                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
            )

        if encoder_hid_dim_type == "text_proj":
            self.encoder_hid_proj = nn.Dense(encoder_hid_dim, cross_attention_dim)
        elif encoder_hid_dim_type == "text_image_proj":
            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
            self.encoder_hid_proj = TextImageProjection(
                text_embed_dim=encoder_hid_dim,
                image_embed_dim=cross_attention_dim,
                cross_attention_dim=cross_attention_dim,
            )
        elif encoder_hid_dim_type == "image_proj":
            # Kandinsky 2.2
            self.encoder_hid_proj = ImageProjection(
                image_embed_dim=encoder_hid_dim,
                cross_attention_dim=cross_attention_dim,
            )
        elif encoder_hid_dim_type is not None:
            raise ValueError(
                f"encoder_hid_dim_type: {encoder_hid_dim_type} must be None, 'text_proj' or 'text_image_proj'."
            )
        else:
            self.encoder_hid_proj = None

    def _set_class_embedding(
        self,
        class_embed_type: Optional[str],
        act_fn: str,
        num_class_embeds: Optional[int],
        projection_class_embeddings_input_dim: Optional[int],
        time_embed_dim: int,
        timestep_input_dim: int,
    ):
        if class_embed_type is None and num_class_embeds is not None:
            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
        elif class_embed_type == "timestep":
            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
        elif class_embed_type == "identity":
            self.class_embedding = nn.Identity()
        elif class_embed_type == "projection":
            if projection_class_embeddings_input_dim is None:
                raise ValueError(
                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
                )
            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
            # 2. it projects from an arbitrary input dimension.
            #
            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
        elif class_embed_type == "simple_projection":
            if projection_class_embeddings_input_dim is None:
                raise ValueError(
                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
                )
            self.class_embedding = nn.Dense(projection_class_embeddings_input_dim, time_embed_dim)
        else:
            self.class_embedding = None

    def _set_add_embedding(
        self,
        addition_embed_type: str,
        addition_embed_type_num_heads: int,
        addition_time_embed_dim: Optional[int],
        flip_sin_to_cos: bool,
        freq_shift: float,
        cross_attention_dim: Optional[int],
        encoder_hid_dim: Optional[int],
        projection_class_embeddings_input_dim: Optional[int],
        time_embed_dim: int,
    ):
        if addition_embed_type == "text":
            if encoder_hid_dim is not None:
                text_time_embedding_from_dim = encoder_hid_dim
            else:
                text_time_embedding_from_dim = cross_attention_dim

            self.add_embedding = TextTimeEmbedding(
                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
            )
        elif addition_embed_type == "text_image":
            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
            self.add_embedding = TextImageTimeEmbedding(
                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
            )
        elif addition_embed_type == "text_time":
            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
        elif addition_embed_type == "image":
            # Kandinsky 2.2
            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
        elif addition_embed_type == "image_hint":
            # Kandinsky 2.2 ControlNet
            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
        elif addition_embed_type is not None:
            raise ValueError(f"addition_embed_type: {addition_embed_type} must be None, 'text' or 'text_image'.")

    def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
        if attention_type in ["gated", "gated-text-image"]:
            positive_len = 768
            if isinstance(cross_attention_dim, int):
                positive_len = cross_attention_dim
            elif isinstance(cross_attention_dim, (list, tuple)):
                positive_len = cross_attention_dim[0]

            feature_type = "text-only" if attention_type == "gated" else "text-image"
            self.position_net = GLIGENTextBoundingboxProjection(
                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
            )

    @property
    def attn_processors(self) -> Dict[str, AttentionProcessor]:
        r"""
        Returns:
            `dict` of attention processors: A dictionary containing all attention processors used in the model with
            indexed by its weight name.
        """
        # set recursively
        processors = {}

        def fn_recursive_add_processors(name: str, module: nn.Cell, processors: Dict[str, AttentionProcessor]):
            if hasattr(module, "get_processor"):
                processors[f"{name}.processor"] = module.get_processor()

            for sub_name, child in module.name_cells().items():
                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)

            return processors

        for name, module in self.name_cells().items():
            fn_recursive_add_processors(name, module, processors)

        return processors

    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        r"""
        Sets the attention processor to use to compute attention.

        Parameters:
            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
                The instantiated processor class or a dictionary of processor classes that will be set as the processor
                for **all** `Attention` layers.

                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
                processor. This is strongly recommended when setting trainable attention processors.

        """
        count = len(self.attn_processors.keys())

        if isinstance(processor, dict) and len(processor) != count:
            raise ValueError(
                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
            )

        def fn_recursive_attn_processor(name: str, module: nn.Cell, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor)
                else:
                    module.set_processor(processor.pop(f"{name}.processor"))

            for sub_name, child in module.name_cells().items():
                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

        for name, module in self.name_cells().items():
            fn_recursive_attn_processor(name, module, processor)

    def set_default_attn_processor(self):
        """
        Disables custom attention processors and sets the default attention implementation.
        """
        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
            processor = AttnAddedKVProcessor()
        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
            processor = AttnProcessor()
        else:
            raise ValueError(
                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
            )

        self.set_attn_processor(processor)

    def _set_gradient_checkpointing(self, module, value=False):
        if hasattr(module, "gradient_checkpointing"):
            module.gradient_checkpointing = value

    def get_time_embed(self, sample: ms.Tensor, timestep: Union[ms.Tensor, float, int]) -> Optional[ms.Tensor]:
        timesteps = timestep
        if not ops.is_tensor(timesteps):
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
            if isinstance(timestep, float):
                dtype = ms.float64
            else:
                dtype = ms.int64
            timesteps = ms.tensor([timesteps], dtype=dtype)
        elif len(timesteps.shape) == 0:
            timesteps = timesteps[None]

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        if timesteps.shape[0] == 1:
            timesteps = timesteps.tile((sample.shape[0],))

        t_emb = self.time_proj(timesteps)

        # `Timesteps` does not contain any weights and will always return f32 tensors
        # but time_embedding might actually be running in fp16. so we need to cast here.
        # there might be better ways to encapsulate this.
        t_emb = t_emb.to(dtype=sample.dtype)
        return t_emb

    def get_class_embed(self, sample: ms.Tensor, class_labels: Optional[ms.Tensor]) -> Optional[ms.Tensor]:
        class_emb = None
        if self.class_embedding is not None:
            if class_labels is None:
                raise ValueError("class_labels should be provided when num_class_embeds > 0")

            if self.class_embed_type == "timestep":
                class_labels = self.time_proj(class_labels)

                # `Timesteps` does not contain any weights and will always return f32 tensors
                # there might be better ways to encapsulate this.
                class_labels = class_labels.to(dtype=sample.dtype)

            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
        return class_emb

    def get_aug_embed(
        self, emb: ms.Tensor, encoder_hidden_states: ms.Tensor, added_cond_kwargs: Dict[str, Any]
    ) -> Optional[ms.Tensor]:
        aug_emb = None
        if self.addition_embed_type == "text":
            aug_emb = self.add_embedding(encoder_hidden_states)
        elif self.addition_embed_type == "text_image":
            # Kandinsky 2.1 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"  # noqa: E501
                )

            image_embs = added_cond_kwargs.get("image_embeds")
            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
            aug_emb = self.add_embedding(text_embs, image_embs)
        elif self.addition_embed_type == "text_time":
            # SDXL - style
            if "text_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"  # noqa: E501
                )
            text_embeds = added_cond_kwargs.get("text_embeds")
            if "time_ids" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"  # noqa: E501
                )
            time_ids = added_cond_kwargs.get("time_ids")
            time_embeds = self.add_time_proj(time_ids.flatten())
            # `Timesteps` does not contain any weights and will always return f32 tensors
            # there might be better ways to encapsulate this.
            time_embeds = time_embeds.to(emb.dtype)
            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
            add_embeds = ops.concat([text_embeds, time_embeds], axis=-1)
            add_embeds = add_embeds.to(emb.dtype)
            aug_emb = self.add_embedding(add_embeds)
        elif self.addition_embed_type == "image":
            # Kandinsky 2.2 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"  # noqa: E501
                )
            image_embs = added_cond_kwargs.get("image_embeds")
            aug_emb = self.add_embedding(image_embs)
        elif self.addition_embed_type == "image_hint":
            # Kandinsky 2.2 - style
            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"  # noqa: E501
                )
            image_embs = added_cond_kwargs.get("image_embeds")
            hint = added_cond_kwargs.get("hint")
            aug_emb = self.add_embedding(image_embs, hint)
        return aug_emb

    def process_encoder_hidden_states(
        self, encoder_hidden_states: ms.Tensor, added_cond_kwargs: Dict[str, Any]
    ) -> ms.Tensor:
        if self.encoder_hid_proj is not None and self.encoder_hid_dim_type == "text_proj":
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
        elif self.encoder_hid_proj is not None and self.encoder_hid_dim_type == "text_image_proj":
            # Kandinsky 2.1 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"  # noqa: E501
                )

            image_embeds = added_cond_kwargs.get("image_embeds")
            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
        elif self.encoder_hid_proj is not None and self.encoder_hid_dim_type == "image_proj":
            # Kandinsky 2.2 - style
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"  # noqa: E501
                )
            image_embeds = added_cond_kwargs.get("image_embeds")
            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
        elif self.encoder_hid_proj is not None and self.encoder_hid_dim_type == "ip_image_proj":
            if "image_embeds" not in added_cond_kwargs:
                raise ValueError(
                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"  # noqa: E501
                )

            if (
                self.has_text_encoder_hid_proj
            ):  # abandon `hasattr` as it is not supported by static graph before MindSpore 2.3
                encoder_hidden_states = self.text_encoder_hid_proj(encoder_hidden_states)

            image_embeds = added_cond_kwargs.get("image_embeds")
            image_embeds = self.encoder_hid_proj(image_embeds)
            encoder_hidden_states = (encoder_hidden_states, image_embeds)
        return encoder_hidden_states

    def construct(
        self,
        sample: ms.Tensor,
        timestep: Union[ms.Tensor, float, int],
        encoder_hidden_states: ms.Tensor,
        class_labels: Optional[ms.Tensor] = None,
        timestep_cond: Optional[ms.Tensor] = None,
        attention_mask: Optional[ms.Tensor] = None,
        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
        added_cond_kwargs: Optional[Dict[str, ms.Tensor]] = None,
        down_block_additional_residuals: Optional[Tuple[ms.Tensor]] = None,
        mid_block_additional_residual: Optional[ms.Tensor] = None,
        down_intrablock_additional_residuals: Optional[Tuple[ms.Tensor]] = None,
        encoder_attention_mask: Optional[ms.Tensor] = None,
        return_dict: bool = False,
    ) -> Union[UNet2DConditionOutput, Tuple]:
        r"""
        The [`UNet2DConditionModel`] forward method.

        Args:
            sample (`ms.Tensor`):
                The noisy input tensor with the following shape `(batch, channel, height, width)`.
            timestep (`ms.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
            encoder_hidden_states (`ms.Tensor`):
                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
            class_labels (`ms.Tensor`, *optional*, defaults to `None`):
                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
            timestep_cond: (`ms.Tensor`, *optional*, defaults to `None`):
                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
                through the `self.time_embedding` layer to obtain the timestep embeddings.
            attention_mask (`ms.Tensor`, *optional*, defaults to `None`):
                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
                negative values to the attention scores corresponding to "discard" tokens.
            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).
            added_cond_kwargs: (`dict`, *optional*):
                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
                are passed along to the UNet blocks.
            down_block_additional_residuals: (`tuple` of `ms.Tensor`, *optional*):
                A tuple of tensors that if specified are added to the residuals of down unet blocks.
            mid_block_additional_residual: (`ms.Tensor`, *optional*):
                A tensor that if specified is added to the residual of the middle unet block.
            down_intrablock_additional_residuals (`tuple` of `ms.Tensor`, *optional*):
                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
            encoder_attention_mask (`ms.Tensor`):
                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
                which adds large negative values to the attention scores corresponding to "discard" tokens.
            return_dict (`bool`, *optional*, defaults to `False`):
                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.

        Returns:
            [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
                If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned,
                otherwise a `tuple` is returned where the first element is the sample tensor.
        """
        # By default samples have to be AT least a multiple of the overall upsampling factor.
        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
        # However, the upsampling interpolation output size can be forced to fit any upsampling size
        # on the fly if necessary.
        default_overall_up_factor = 2**self.num_upsamplers

        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
        forward_upsample_size = False
        upsample_size = None

        for dim in sample.shape[-2:]:
            if dim % default_overall_up_factor != 0:
                # Forward upsample size to force interpolation output size.
                forward_upsample_size = True
                break

        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
        # expects mask of shape:
        #   [batch, key_tokens]
        # adds singleton query_tokens dimension:
        #   [batch,                    1, key_tokens]
        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
        if attention_mask is not None:
            # assume that mask is expressed as:
            #   (1 = keep,      0 = discard)
            # convert mask into a bias that can be added to attention scores:
            #       (keep = +0,     discard = -10000.0)
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

        # convert encoder_attention_mask to a bias the same way we do for attention_mask
        if encoder_attention_mask is not None:
            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)

        # 0. center input if necessary
        if self.center_input_sample:
            sample = 2 * sample - 1.0

        # 1. time
        t_emb = self.get_time_embed(sample=sample, timestep=timestep)
        emb = self.time_embedding(t_emb, timestep_cond)
        aug_emb = None

        class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
        if class_emb is not None:
            if self.class_embeddings_concat:
                emb = ops.cat([emb, class_emb], axis=-1)
            else:
                emb = emb + class_emb

        aug_emb = self.get_aug_embed(
            emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
        )
        if self.addition_embed_type == "image_hint":
            aug_emb, hint = aug_emb
            sample = ops.cat([sample, hint], axis=1)

        emb = emb + aug_emb if aug_emb is not None else emb

        if self.time_embed_act is not None:
            emb = self.time_embed_act(emb)

        encoder_hidden_states = self.process_encoder_hidden_states(
            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
        )

        # 2. pre-process
        sample = self.conv_in(sample)

        # 2.5 GLIGEN position net
        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
            # copy and pop cross_attention_kwargs manually, as dict.copy/pop is not supported in GRAPH MODE syntax
            copied_cross_attention_kwargs = {}
            for k, v in cross_attention_kwargs.items():
                if k == "gligen":
                    copied_cross_attention_kwargs[k] = {"objs": self.position_net(**v)}
                else:
                    copied_cross_attention_kwargs[k] = v
            cross_attention_kwargs = copied_cross_attention_kwargs

        # 3. down
        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
        if cross_attention_kwargs is not None and "scale" in cross_attention_kwargs:
            # weight the lora layers by setting `lora_scale` for each PEFT layer here
            # and remove `lora_scale` from each PEFT layer at the end.
            # scale_lora_layers & unscale_lora_layers maybe contains some operation forbidden in graph mode
            raise RuntimeError(
                f"You are trying to set scaling of lora layer by passing {cross_attention_kwargs['scale']=}. "
                f"However it's not allowed in on-the-fly model forwarding. "
                f"Please manually call `scale_lora_layers(model, lora_scale)` before model forwarding and "
                f"`unscale_lora_layers(model, lora_scale)` after model forwarding. "
                f"For example, it can be done in a pipeline call like `StableDiffusionPipeline.__call__`."
            )

        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
        is_adapter = down_intrablock_additional_residuals is not None
        # using variable `adapter_index` to get item in `down_intrablock_additional_residuals` for avoiding
        # pop operations in construct(), which are not fully supported in GRAPH_MODE
        adapter_index = 0
        # maintain backward compatibility for legacy usage, where
        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
        #       but can only use one or the other
        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
            # deprecate(
            #     "T2I should not use down_block_additional_residuals",
            #     "1.3.0",
            #     "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
            #            and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
            #            for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
            #     standard_warn=False,
            # )
            down_intrablock_additional_residuals = down_block_additional_residuals
            is_adapter = True

        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            if downsample_block.has_cross_attention:
                # For t2i-adapter CrossAttnDownBlock2D
                additional_residuals = {}
                if is_adapter and len(down_intrablock_additional_residuals) > adapter_index:
                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals[adapter_index]
                    adapter_index += 1

                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
                    **additional_residuals,
                )
            else:
                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
                if is_adapter and len(down_intrablock_additional_residuals) > adapter_index:
                    # `sample` here is one of element in `res_samples`, in PyTorch they refer to the same object
                    # which means changes on sample will take effect on the counterpart in res_samples. However it
                    # doesn't work in MindSpore as they are different objects thus we need change both of them manually.
                    sample += down_intrablock_additional_residuals[adapter_index]
                    res_samples = list(res_samples)  # convert to list to support item assignment
                    res_samples[-1] += down_intrablock_additional_residuals[adapter_index]
                    res_samples = tuple(res_samples)  # convert back to tuple to concat
                    adapter_index += 1

            down_block_res_samples += res_samples

        if is_controlnet:
            new_down_block_res_samples = ()

            for down_block_res_sample, down_block_additional_residual in zip(
                down_block_res_samples, down_block_additional_residuals
            ):
                down_block_res_sample = down_block_res_sample + down_block_additional_residual
                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)

            down_block_res_samples = new_down_block_res_samples

        # 4. mid
        if self.mid_block is not None:
            if self.mid_block.has_cross_attention:
                sample = self.mid_block(
                    sample,
                    emb,
                    encoder_hidden_states=encoder_hidden_states,
                    attention_mask=attention_mask,
                    cross_attention_kwargs=cross_attention_kwargs,
                    encoder_attention_mask=encoder_attention_mask,
                )
            else:
                sample = self.mid_block(sample, emb)

            # To support T2I-Adapter-XL
            if (
                is_adapter
                and len(down_intrablock_additional_residuals) > adapter_index
                and sample.shape == down_intrablock_additional_residuals[adapter_index].shape
            ):
                sample += down_intrablock_additional_residuals[adapter_index]
                adapter_index += 1

        if is_controlnet:
            sample = sample + mid_block_additional_residual

        # 5. up
        for i, upsample_block in enumerate(self.up_blocks):
            is_final_block = i == len(self.up_blocks) - 1

            res_samples = down_block_res_samples[-self.layers_per_resnet_in_up_blocks[i] :]
            down_block_res_samples = down_block_res_samples[: -self.layers_per_resnet_in_up_blocks[i]]

            # if we have not reached the final block and need to forward the
            # upsample size, we do it here
            if not is_final_block and forward_upsample_size:
                upsample_size = down_block_res_samples[-1].shape[2:]

            if upsample_block.has_cross_attention:
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    encoder_hidden_states=encoder_hidden_states,
                    cross_attention_kwargs=cross_attention_kwargs,
                    upsample_size=upsample_size,
                    attention_mask=attention_mask,
                    encoder_attention_mask=encoder_attention_mask,
                )
            else:
                sample = upsample_block(
                    hidden_states=sample,
                    temb=emb,
                    res_hidden_states_tuple=res_samples,
                    upsample_size=upsample_size,
                )

        # 6. post-process
        if self.conv_norm_out:
            sample = self.conv_norm_out(sample)
            sample = self.conv_act(sample)
        sample = self.conv_out(sample)

        if not return_dict:
            return (sample,)

        return UNet2DConditionOutput(sample=sample)

mindone.diffusers.UNet2DConditionModel.attn_processors: Dict[str, AttentionProcessor] property

RETURNS	DESCRIPTION
Dict[str, AttentionProcessor]	dict of attention processors: A dictionary containing all attention processors used in the model with
Dict[str, AttentionProcessor]	indexed by its weight name.

mindone.diffusers.UNet2DConditionModel.construct(sample, timestep, encoder_hidden_states, class_labels=None, timestep_cond=None, attention_mask=None, cross_attention_kwargs=None, added_cond_kwargs=None, down_block_additional_residuals=None, mid_block_additional_residual=None, down_intrablock_additional_residuals=None, encoder_attention_mask=None, return_dict=False)

The [UNet2DConditionModel] forward method.

PARAMETER	DESCRIPTION
sample	The noisy input tensor with the following shape (batch, channel, height, width). TYPE: `ms.Tensor`
timestep	The number of timesteps to denoise an input. TYPE: `ms.Tensor` or `float` or `int`
encoder_hidden_states	The encoder hidden states with shape (batch, sequence_length, feature_dim). TYPE: `ms.Tensor`
class_labels	Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings. TYPE: `ms.Tensor`, *optional*, defaults to `None` DEFAULT: None
timestep_cond	(ms.Tensor, optional, defaults to None): Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed through the self.time_embedding layer to obtain the timestep embeddings. TYPE: Optional[Tensor] DEFAULT: None
attention_mask	An attention mask of shape (batch, key_tokens) is applied to encoder_hidden_states. If 1 the mask is kept, otherwise if 0 it is discarded. Mask will be converted into a bias, which adds large negative values to the attention scores corresponding to "discard" tokens. TYPE: `ms.Tensor`, *optional*, defaults to `None` DEFAULT: None
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
added_cond_kwargs	(dict, optional): A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that are passed along to the UNet blocks. TYPE: Optional[Dict[str, Tensor]] DEFAULT: None
down_block_additional_residuals	(tuple of ms.Tensor, optional): A tuple of tensors that if specified are added to the residuals of down unet blocks. TYPE: Optional[Tuple[Tensor]] DEFAULT: None
mid_block_additional_residual	(ms.Tensor, optional): A tensor that if specified is added to the residual of the middle unet block. TYPE: Optional[Tensor] DEFAULT: None
down_intrablock_additional_residuals	additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s) TYPE: `tuple` of `ms.Tensor`, *optional* DEFAULT: None
encoder_attention_mask	A cross-attention mask of shape (batch, sequence_length) is applied to encoder_hidden_states. If True the mask is kept, otherwise if False it is discarded. Mask will be converted into a bias, which adds large negative values to the attention scores corresponding to "discard" tokens. TYPE: `ms.Tensor` DEFAULT: None
return_dict	Whether or not to return a [~models.unets.unet_2d_condition.UNet2DConditionOutput] instead of a plain tuple. TYPE: `bool`, *optional*, defaults to `False` DEFAULT: False

RETURNS	DESCRIPTION
Union[UNet2DConditionOutput, Tuple]	[~models.unets.unet_2d_condition.UNet2DConditionOutput] or tuple: If return_dict is True, an [~models.unets.unet_2d_condition.UNet2DConditionOutput] is returned, otherwise a tuple is returned where the first element is the sample tensor.

Source code in mindone/diffusers/models/unets/unet_2d_condition.py

def construct(
    self,
    sample: ms.Tensor,
    timestep: Union[ms.Tensor, float, int],
    encoder_hidden_states: ms.Tensor,
    class_labels: Optional[ms.Tensor] = None,
    timestep_cond: Optional[ms.Tensor] = None,
    attention_mask: Optional[ms.Tensor] = None,
    cross_attention_kwargs: Optional[Dict[str, Any]] = None,
    added_cond_kwargs: Optional[Dict[str, ms.Tensor]] = None,
    down_block_additional_residuals: Optional[Tuple[ms.Tensor]] = None,
    mid_block_additional_residual: Optional[ms.Tensor] = None,
    down_intrablock_additional_residuals: Optional[Tuple[ms.Tensor]] = None,
    encoder_attention_mask: Optional[ms.Tensor] = None,
    return_dict: bool = False,
) -> Union[UNet2DConditionOutput, Tuple]:
    r"""
    The [`UNet2DConditionModel`] forward method.

    Args:
        sample (`ms.Tensor`):
            The noisy input tensor with the following shape `(batch, channel, height, width)`.
        timestep (`ms.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
        encoder_hidden_states (`ms.Tensor`):
            The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
        class_labels (`ms.Tensor`, *optional*, defaults to `None`):
            Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
        timestep_cond: (`ms.Tensor`, *optional*, defaults to `None`):
            Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
            through the `self.time_embedding` layer to obtain the timestep embeddings.
        attention_mask (`ms.Tensor`, *optional*, defaults to `None`):
            An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
            is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
            negative values to the attention scores corresponding to "discard" tokens.
        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).
        added_cond_kwargs: (`dict`, *optional*):
            A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
            are passed along to the UNet blocks.
        down_block_additional_residuals: (`tuple` of `ms.Tensor`, *optional*):
            A tuple of tensors that if specified are added to the residuals of down unet blocks.
        mid_block_additional_residual: (`ms.Tensor`, *optional*):
            A tensor that if specified is added to the residual of the middle unet block.
        down_intrablock_additional_residuals (`tuple` of `ms.Tensor`, *optional*):
            additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
        encoder_attention_mask (`ms.Tensor`):
            A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
            `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
            which adds large negative values to the attention scores corresponding to "discard" tokens.
        return_dict (`bool`, *optional*, defaults to `False`):
            Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
            tuple.

    Returns:
        [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
            If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned,
            otherwise a `tuple` is returned where the first element is the sample tensor.
    """
    # By default samples have to be AT least a multiple of the overall upsampling factor.
    # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
    # However, the upsampling interpolation output size can be forced to fit any upsampling size
    # on the fly if necessary.
    default_overall_up_factor = 2**self.num_upsamplers

    # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
    forward_upsample_size = False
    upsample_size = None

    for dim in sample.shape[-2:]:
        if dim % default_overall_up_factor != 0:
            # Forward upsample size to force interpolation output size.
            forward_upsample_size = True
            break

    # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
    # expects mask of shape:
    #   [batch, key_tokens]
    # adds singleton query_tokens dimension:
    #   [batch,                    1, key_tokens]
    # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
    #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
    #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
    if attention_mask is not None:
        # assume that mask is expressed as:
        #   (1 = keep,      0 = discard)
        # convert mask into a bias that can be added to attention scores:
        #       (keep = +0,     discard = -10000.0)
        attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
        attention_mask = attention_mask.unsqueeze(1)

    # convert encoder_attention_mask to a bias the same way we do for attention_mask
    if encoder_attention_mask is not None:
        encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
        encoder_attention_mask = encoder_attention_mask.unsqueeze(1)

    # 0. center input if necessary
    if self.center_input_sample:
        sample = 2 * sample - 1.0

    # 1. time
    t_emb = self.get_time_embed(sample=sample, timestep=timestep)
    emb = self.time_embedding(t_emb, timestep_cond)
    aug_emb = None

    class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
    if class_emb is not None:
        if self.class_embeddings_concat:
            emb = ops.cat([emb, class_emb], axis=-1)
        else:
            emb = emb + class_emb

    aug_emb = self.get_aug_embed(
        emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
    )
    if self.addition_embed_type == "image_hint":
        aug_emb, hint = aug_emb
        sample = ops.cat([sample, hint], axis=1)

    emb = emb + aug_emb if aug_emb is not None else emb

    if self.time_embed_act is not None:
        emb = self.time_embed_act(emb)

    encoder_hidden_states = self.process_encoder_hidden_states(
        encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
    )

    # 2. pre-process
    sample = self.conv_in(sample)

    # 2.5 GLIGEN position net
    if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
        # copy and pop cross_attention_kwargs manually, as dict.copy/pop is not supported in GRAPH MODE syntax
        copied_cross_attention_kwargs = {}
        for k, v in cross_attention_kwargs.items():
            if k == "gligen":
                copied_cross_attention_kwargs[k] = {"objs": self.position_net(**v)}
            else:
                copied_cross_attention_kwargs[k] = v
        cross_attention_kwargs = copied_cross_attention_kwargs

    # 3. down
    # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
    # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
    if cross_attention_kwargs is not None and "scale" in cross_attention_kwargs:
        # weight the lora layers by setting `lora_scale` for each PEFT layer here
        # and remove `lora_scale` from each PEFT layer at the end.
        # scale_lora_layers & unscale_lora_layers maybe contains some operation forbidden in graph mode
        raise RuntimeError(
            f"You are trying to set scaling of lora layer by passing {cross_attention_kwargs['scale']=}. "
            f"However it's not allowed in on-the-fly model forwarding. "
            f"Please manually call `scale_lora_layers(model, lora_scale)` before model forwarding and "
            f"`unscale_lora_layers(model, lora_scale)` after model forwarding. "
            f"For example, it can be done in a pipeline call like `StableDiffusionPipeline.__call__`."
        )

    is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
    # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
    is_adapter = down_intrablock_additional_residuals is not None
    # using variable `adapter_index` to get item in `down_intrablock_additional_residuals` for avoiding
    # pop operations in construct(), which are not fully supported in GRAPH_MODE
    adapter_index = 0
    # maintain backward compatibility for legacy usage, where
    #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
    #       but can only use one or the other
    if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
        # deprecate(
        #     "T2I should not use down_block_additional_residuals",
        #     "1.3.0",
        #     "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
        #            and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
        #            for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
        #     standard_warn=False,
        # )
        down_intrablock_additional_residuals = down_block_additional_residuals
        is_adapter = True

    down_block_res_samples = (sample,)
    for downsample_block in self.down_blocks:
        if downsample_block.has_cross_attention:
            # For t2i-adapter CrossAttnDownBlock2D
            additional_residuals = {}
            if is_adapter and len(down_intrablock_additional_residuals) > adapter_index:
                additional_residuals["additional_residuals"] = down_intrablock_additional_residuals[adapter_index]
                adapter_index += 1

            sample, res_samples = downsample_block(
                hidden_states=sample,
                temb=emb,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                cross_attention_kwargs=cross_attention_kwargs,
                encoder_attention_mask=encoder_attention_mask,
                **additional_residuals,
            )
        else:
            sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
            if is_adapter and len(down_intrablock_additional_residuals) > adapter_index:
                # `sample` here is one of element in `res_samples`, in PyTorch they refer to the same object
                # which means changes on sample will take effect on the counterpart in res_samples. However it
                # doesn't work in MindSpore as they are different objects thus we need change both of them manually.
                sample += down_intrablock_additional_residuals[adapter_index]
                res_samples = list(res_samples)  # convert to list to support item assignment
                res_samples[-1] += down_intrablock_additional_residuals[adapter_index]
                res_samples = tuple(res_samples)  # convert back to tuple to concat
                adapter_index += 1

        down_block_res_samples += res_samples

    if is_controlnet:
        new_down_block_res_samples = ()

        for down_block_res_sample, down_block_additional_residual in zip(
            down_block_res_samples, down_block_additional_residuals
        ):
            down_block_res_sample = down_block_res_sample + down_block_additional_residual
            new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)

        down_block_res_samples = new_down_block_res_samples

    # 4. mid
    if self.mid_block is not None:
        if self.mid_block.has_cross_attention:
            sample = self.mid_block(
                sample,
                emb,
                encoder_hidden_states=encoder_hidden_states,
                attention_mask=attention_mask,
                cross_attention_kwargs=cross_attention_kwargs,
                encoder_attention_mask=encoder_attention_mask,
            )
        else:
            sample = self.mid_block(sample, emb)

        # To support T2I-Adapter-XL
        if (
            is_adapter
            and len(down_intrablock_additional_residuals) > adapter_index
            and sample.shape == down_intrablock_additional_residuals[adapter_index].shape
        ):
            sample += down_intrablock_additional_residuals[adapter_index]
            adapter_index += 1

    if is_controlnet:
        sample = sample + mid_block_additional_residual

    # 5. up
    for i, upsample_block in enumerate(self.up_blocks):
        is_final_block = i == len(self.up_blocks) - 1

        res_samples = down_block_res_samples[-self.layers_per_resnet_in_up_blocks[i] :]
        down_block_res_samples = down_block_res_samples[: -self.layers_per_resnet_in_up_blocks[i]]

        # if we have not reached the final block and need to forward the
        # upsample size, we do it here
        if not is_final_block and forward_upsample_size:
            upsample_size = down_block_res_samples[-1].shape[2:]

        if upsample_block.has_cross_attention:
            sample = upsample_block(
                hidden_states=sample,
                temb=emb,
                res_hidden_states_tuple=res_samples,
                encoder_hidden_states=encoder_hidden_states,
                cross_attention_kwargs=cross_attention_kwargs,
                upsample_size=upsample_size,
                attention_mask=attention_mask,
                encoder_attention_mask=encoder_attention_mask,
            )
        else:
            sample = upsample_block(
                hidden_states=sample,
                temb=emb,
                res_hidden_states_tuple=res_samples,
                upsample_size=upsample_size,
            )

    # 6. post-process
    if self.conv_norm_out:
        sample = self.conv_norm_out(sample)
        sample = self.conv_act(sample)
    sample = self.conv_out(sample)

    if not return_dict:
        return (sample,)

    return UNet2DConditionOutput(sample=sample)

mindone.diffusers.UNet2DConditionModel.set_attn_processor(processor)

Sets the attention processor to use to compute attention.

PARAMETER	DESCRIPTION
processor	The instantiated processor class or a dictionary of processor classes that will be set as the processor for all Attention layers. If processor is a dict, the key needs to define the path to the corresponding cross attention processor. This is strongly recommended when setting trainable attention processors. TYPE: `dict` of `AttentionProcessor` or only `AttentionProcessor`

Source code in mindone/diffusers/models/unets/unet_2d_condition.py

def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
    r"""
    Sets the attention processor to use to compute attention.

    Parameters:
        processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
            The instantiated processor class or a dictionary of processor classes that will be set as the processor
            for **all** `Attention` layers.

            If `processor` is a dict, the key needs to define the path to the corresponding cross attention
            processor. This is strongly recommended when setting trainable attention processors.

    """
    count = len(self.attn_processors.keys())

    if isinstance(processor, dict) and len(processor) != count:
        raise ValueError(
            f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
            f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
        )

    def fn_recursive_attn_processor(name: str, module: nn.Cell, processor):
        if hasattr(module, "set_processor"):
            if not isinstance(processor, dict):
                module.set_processor(processor)
            else:
                module.set_processor(processor.pop(f"{name}.processor"))

        for sub_name, child in module.name_cells().items():
            fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)

    for name, module in self.name_cells().items():
        fn_recursive_attn_processor(name, module, processor)

mindone.diffusers.UNet2DConditionModel.set_default_attn_processor()

Disables custom attention processors and sets the default attention implementation.

Source code in mindone/diffusers/models/unets/unet_2d_condition.py

def set_default_attn_processor(self):
    """
    Disables custom attention processors and sets the default attention implementation.
    """
    if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
        processor = AttnAddedKVProcessor()
    elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
        processor = AttnProcessor()
    else:
        raise ValueError(
            f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
        )

    self.set_attn_processor(processor)

mindone.diffusers.models.unets.unet_2d_condition.UNet2DConditionOutput dataclass

Bases: BaseOutput

The output of [UNet2DConditionModel].

PARAMETER	DESCRIPTION
sample	The hidden states output conditioned on encoder_hidden_states input. Output of last layer of model. TYPE: `ms.Tensor` of shape `(batch_size, num_channels, height, width)` DEFAULT: None

Source code in mindone/diffusers/models/unets/unet_2d_condition.py

@dataclass
class UNet2DConditionOutput(BaseOutput):
    """
    The output of [`UNet2DConditionModel`].

    Args:
        sample (`ms.Tensor` of shape `(batch_size, num_channels, height, width)`):
            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
    """

    sample: ms.Tensor = None