models/unet_2d_condition.py

# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_condition.py
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.utils.checkpoint
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import UNet2DConditionLoadersMixin
from diffusers.models.activations import get_activation
from diffusers.models.attention_processor import (
    ADDED_KV_ATTENTION_PROCESSORS,
    CROSS_ATTENTION_PROCESSORS,
    AttentionProcessor,
    AttnAddedKVProcessor,
    AttnProcessor,
)
from diffusers.models.embeddings import (
    GaussianFourierProjection,
    ImageHintTimeEmbedding,
    ImageProjection,
    ImageTimeEmbedding,
    TextImageProjection,
    TextImageTimeEmbedding,
    TextTimeEmbedding,
    TimestepEmbedding,
    Timesteps,
)
from diffusers.models.modeling_utils import ModelMixin
from diffusers.utils import (
    USE_PEFT_BACKEND,
    BaseOutput,
    deprecate,
    logging,
    scale_lora_layers,
    unscale_lora_layers,
)

from .unet_2d_blocks import (
    UNetMidBlock2D,
    UNetMidBlock2DCrossAttn,
    get_down_block,
    get_up_block,
)

logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


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

    Args:
        sample (`torch.FloatTensor` 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: torch.FloatTensor = None
    ref_features: Tuple[torch.FloatTensor] = None


class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin):
    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 `False`):
            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.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
            [`~models.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.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
            [`~models.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

    @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: int = 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=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."
            )

        # 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
        # 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
        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}."
            )

        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}."
            )

        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}."
            )

        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}."
            )

        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}."
            )

        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}."
            )

        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}."
            )
        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."
                    )

        # 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,
            padding=conv_in_padding,
        )

        # time
        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`."
            )

        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,
        )

        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.Linear(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"` (Kadinsky 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

        # class embedding
        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(time_embed_dim, time_embed_dim)
        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.Linear(
                projection_class_embeddings_input_dim, time_embed_dim
            )
        else:
            self.class_embedding = None

        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"` (Kadinsky 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'."
            )

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

        self.down_blocks = nn.ModuleList([])
        self.up_blocks = nn.ModuleList([])

        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

        # down
        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,
            )
            self.down_blocks.append(down_block)

        # mid
        if mid_block_type == "UNetMidBlock2DCrossAttn":
            self.mid_block = UNetMidBlock2DCrossAttn(
                transformer_layers_per_block=transformer_layers_per_block[-1],
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                dropout=dropout,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_time_scale_shift=resnet_time_scale_shift,
                cross_attention_dim=cross_attention_dim[-1],
                num_attention_heads=num_attention_heads[-1],
                resnet_groups=norm_num_groups,
                dual_cross_attention=dual_cross_attention,
                use_linear_projection=use_linear_projection,
                upcast_attention=upcast_attention,
                attention_type=attention_type,
            )
        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
            raise NotImplementedError(f"Unsupport mid_block_type: {mid_block_type}")
        elif mid_block_type == "UNetMidBlock2D":
            self.mid_block = UNetMidBlock2D(
                in_channels=block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                dropout=dropout,
                num_layers=0,
                resnet_eps=norm_eps,
                resnet_act_fn=act_fn,
                output_scale_factor=mid_block_scale_factor,
                resnet_groups=norm_num_groups,
                resnet_time_scale_shift=resnet_time_scale_shift,
                add_attention=False,
            )
        elif mid_block_type is None:
            self.mid_block = None
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")

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

        # up
        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,
            )
            self.up_blocks.append(up_block)
            prev_output_channel = output_channel

        # out
        if norm_num_groups is not None:
            self.conv_norm_out = nn.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
        self.conv_norm_out = 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,
        #     padding=conv_out_padding,
        # )

        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, tuple) or isinstance(
                cross_attention_dim, list
            ):
                positive_len = cross_attention_dim[0]

            feature_type = "text-only" if attention_type == "gated" else "text-image"
            self.position_net = PositionNet(
                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: torch.nn.Module,
            processors: Dict[str, AttentionProcessor],
        ):
            if hasattr(module, "get_processor"):
                processors[f"{name}.processor"] = module.get_processor(
                    return_deprecated_lora=True
                )

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

            return processors

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

        return processors

    def set_attn_processor(
        self,
        processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]],
        _remove_lora=False,
    ):
        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: torch.nn.Module, processor):
            if hasattr(module, "set_processor"):
                if not isinstance(processor, dict):
                    module.set_processor(processor, _remove_lora=_remove_lora)
                else:
                    module.set_processor(
                        processor.pop(f"{name}.processor"), _remove_lora=_remove_lora
                    )

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

        for name, module in self.named_children():
            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, _remove_lora=True)

    def set_attention_slice(self, slice_size):
        r"""
        Enable sliced attention computation.

        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
        several steps. This is useful for saving some memory in exchange for a small decrease in speed.

        Args:
            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
                must be a multiple of `slice_size`.
        """
        sliceable_head_dims = []

        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
            if hasattr(module, "set_attention_slice"):
                sliceable_head_dims.append(module.sliceable_head_dim)

            for child in module.children():
                fn_recursive_retrieve_sliceable_dims(child)

        # retrieve number of attention layers
        for module in self.children():
            fn_recursive_retrieve_sliceable_dims(module)

        num_sliceable_layers = len(sliceable_head_dims)

        if slice_size == "auto":
            # half the attention head size is usually a good trade-off between
            # speed and memory
            slice_size = [dim // 2 for dim in sliceable_head_dims]
        elif slice_size == "max":
            # make smallest slice possible
            slice_size = num_sliceable_layers * [1]

        slice_size = (
            num_sliceable_layers * [slice_size]
            if not isinstance(slice_size, list)
            else slice_size
        )

        if len(slice_size) != len(sliceable_head_dims):
            raise ValueError(
                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
            )

        for i in range(len(slice_size)):
            size = slice_size[i]
            dim = sliceable_head_dims[i]
            if size is not None and size > dim:
                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")

        # Recursively walk through all the children.
        # Any children which exposes the set_attention_slice method
        # gets the message
        def fn_recursive_set_attention_slice(
            module: torch.nn.Module, slice_size: List[int]
        ):
            if hasattr(module, "set_attention_slice"):
                module.set_attention_slice(slice_size.pop())

            for child in module.children():
                fn_recursive_set_attention_slice(child, slice_size)

        reversed_slice_size = list(reversed(slice_size))
        for module in self.children():
            fn_recursive_set_attention_slice(module, reversed_slice_size)

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

    def enable_freeu(self, s1, s2, b1, b2):
        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.

        The suffixes after the scaling factors represent the stage blocks where they are being applied.

        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.

        Args:
            s1 (`float`):
                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
                mitigate the "oversmoothing effect" in the enhanced denoising process.
            s2 (`float`):
                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
                mitigate the "oversmoothing effect" in the enhanced denoising process.
            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
        """
        for i, upsample_block in enumerate(self.up_blocks):
            setattr(upsample_block, "s1", s1)
            setattr(upsample_block, "s2", s2)
            setattr(upsample_block, "b1", b1)
            setattr(upsample_block, "b2", b2)

    def disable_freeu(self):
        """Disables the FreeU mechanism."""
        freeu_keys = {"s1", "s2", "b1", "b2"}
        for i, upsample_block in enumerate(self.up_blocks):
            for k in freeu_keys:
                if (
                    hasattr(upsample_block, k)
                    or getattr(upsample_block, k, None) is not None
                ):
                    setattr(upsample_block, k, None)

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

        Args:
            sample (`torch.FloatTensor`):
                The noisy input tensor with the following shape `(batch, channel, height, width)`.
            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
            encoder_hidden_states (`torch.FloatTensor`):
                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
            timestep_cond: (`torch.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 (`torch.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 `torch.Tensor`, *optional*):
                A tuple of tensors that if specified are added to the residuals of down unet blocks.
            mid_block_additional_residual: (`torch.Tensor`, *optional*):
                A tensor that if specified is added to the residual of the middle unet block.
            encoder_attention_mask (`torch.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 `True`):
                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
                tuple.
            cross_attention_kwargs (`dict`, *optional*):
                A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
            added_cond_kwargs: (`dict`, *optional*):
                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
                are passed along to the UNet blocks.
            down_block_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
                additional residuals to be added to UNet long skip connections from down blocks to up blocks for
                example from ControlNet side model(s)
            mid_block_additional_residual (`torch.Tensor`, *optional*):
                additional residual to be added to UNet mid block output, for example from ControlNet side model
            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)

        Returns:
            [`~models.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
                If `return_dict` is True, an [`~models.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.config.center_input_sample:
            sample = 2 * sample - 1.0

        # 1. time
        timesteps = timestep
        if not torch.is_tensor(timesteps):
            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
            # This would be a good case for the `match` statement (Python 3.10+)
            is_mps = sample.device.type == "mps"
            if isinstance(timestep, float):
                dtype = torch.float32 if is_mps else torch.float64
            else:
                dtype = torch.int32 if is_mps else torch.int64
            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
        elif len(timesteps.shape) == 0:
            timesteps = timesteps[None].to(sample.device)

        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
        timesteps = timesteps.expand(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)

        emb = self.time_embedding(t_emb, timestep_cond)
        aug_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.config.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)

            if self.config.class_embeddings_concat:
                emb = torch.cat([emb, class_emb], dim=-1)
            else:
                emb = emb + class_emb

        if self.config.addition_embed_type == "text":
            aug_emb = self.add_embedding(encoder_hidden_states)
        elif self.config.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`"
                )

            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.config.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`"
                )
            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`"
                )
            time_ids = added_cond_kwargs.get("time_ids")
            time_embeds = self.add_time_proj(time_ids.flatten())
            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
            add_embeds = add_embeds.to(emb.dtype)
            aug_emb = self.add_embedding(add_embeds)
        elif self.config.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`"
                )
            image_embs = added_cond_kwargs.get("image_embeds")
            aug_emb = self.add_embedding(image_embs)
        elif self.config.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`"
                )
            image_embs = added_cond_kwargs.get("image_embeds")
            hint = added_cond_kwargs.get("hint")
            aug_emb, hint = self.add_embedding(image_embs, hint)
            sample = torch.cat([sample, hint], dim=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)

        if (
            self.encoder_hid_proj is not None
            and self.config.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.config.encoder_hid_dim_type == "text_image_proj"
        ):
            # Kadinsky 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`"
                )

            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.config.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`"
                )
            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.config.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`"
                )
            image_embeds = added_cond_kwargs.get("image_embeds")
            image_embeds = self.encoder_hid_proj(image_embeds).to(
                encoder_hidden_states.dtype
            )
            encoder_hidden_states = torch.cat(
                [encoder_hidden_states, image_embeds], dim=1
            )

        # 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
        ):
            cross_attention_kwargs = cross_attention_kwargs.copy()
            gligen_args = cross_attention_kwargs.pop("gligen")
            cross_attention_kwargs["gligen"] = {
                "objs": self.position_net(**gligen_args)
            }

        # 3. down
        lora_scale = (
            cross_attention_kwargs.get("scale", 1.0)
            if cross_attention_kwargs is not None
            else 1.0
        )
        if USE_PEFT_BACKEND:
            # weight the lora layers by setting `lora_scale` for each PEFT layer
            scale_lora_layers(self, lora_scale)

        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
        # 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,)
        tot_referece_features = ()
        for downsample_block in self.down_blocks:
            if (
                hasattr(downsample_block, "has_cross_attention")
                and downsample_block.has_cross_attention
            ):
                # For t2i-adapter CrossAttnDownBlock2D
                additional_residuals = {}
                if is_adapter and len(down_intrablock_additional_residuals) > 0:
                    additional_residuals[
                        "additional_residuals"
                    ] = down_intrablock_additional_residuals.pop(0)

                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, scale=lora_scale
                )
                if is_adapter and len(down_intrablock_additional_residuals) > 0:
                    sample += down_intrablock_additional_residuals.pop(0)

            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 (
                hasattr(self.mid_block, "has_cross_attention")
                and 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) > 0
                and sample.shape == down_intrablock_additional_residuals[0].shape
            ):
                sample += down_intrablock_additional_residuals.pop(0)

        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[-len(upsample_block.resnets) :]
            down_block_res_samples = down_block_res_samples[
                : -len(upsample_block.resnets)
            ]

            # 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 (
                hasattr(upsample_block, "has_cross_attention")
                and 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,
                    scale=lora_scale,
                )

        # 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 USE_PEFT_BACKEND:
            # remove `lora_scale` from each PEFT layer
            unscale_lora_layers(self, lora_scale)

        if not return_dict:
            return (sample,)

        return UNet2DConditionOutput(sample=sample)