Upload 9 files

models/attention.py

@@ -0,0 +1,777 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+# Some modifications are reimplemented in public environments by Xiao Fu and Mu Hu
+
+
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+import xformers
+
+from diffusers.utils import USE_PEFT_BACKEND
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
+from diffusers.models.attention_processor import Attention
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.lora import LoRACompatibleLinear
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+
+
+def _chunked_feed_forward(
+    ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int, lora_scale: Optional[float] = None
+):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    if lora_scale is None:
+        ff_output = torch.cat(
+            [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+    else:
+        # TOOD(Patrick): LoRA scale can be removed once PEFT refactor is complete
+        ff_output = torch.cat(
+            [ff(hid_slice, scale=lora_scale) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+            dim=chunk_dim,
+        )
+
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+        ada_norm_bias: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if self.use_ada_layer_norm:
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_zero:
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif self.use_ada_layer_norm_continuous:
+            self.norm1 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "rms_norm",
+            )
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+
+        self.attn1 = CustomJointAttention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+
+            if self.use_ada_layer_norm:
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif self.use_ada_layer_norm_continuous:
+                self.norm2 = AdaLayerNormContinuous(
+                    dim,
+                    ada_norm_continous_conditioning_embedding_dim,
+                    norm_elementwise_affine,
+                    norm_eps,
+                    ada_norm_bias,
+                    "rms_norm",
+                )
+            else:
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if self.use_ada_layer_norm_continuous:
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+        elif not self.use_ada_layer_norm_single:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if self.use_ada_layer_norm_single:
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.use_ada_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.use_ada_layer_norm_zero:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.use_layer_norm:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif self.use_ada_layer_norm_single:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 2. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.use_ada_layer_norm_single:
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 2.5 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.use_ada_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.use_ada_layer_norm_zero or self.use_layer_norm:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.use_ada_layer_norm_single:
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.use_ada_layer_norm_continuous:
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.use_ada_layer_norm_single is False:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        if self.use_ada_layer_norm_continuous:
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.use_ada_layer_norm_single:
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(
+                self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size, lora_scale=lora_scale
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states, scale=lora_scale)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.use_ada_layer_norm_single:
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class CustomJointAttention(Attention):
+    def set_use_memory_efficient_attention_xformers(
+        self, use_memory_efficient_attention_xformers: bool, *args, **kwargs
+    ):
+        processor = XFormersJointAttnProcessor()
+        self.set_processor(processor)
+        # print("using xformers attention processor")
+
+
+class XFormersJointAttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __call__(
+        self,
+        attn: Attention,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+        num_tasks=2
+    ):
+    
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        # from yuancheng; here attention_mask is None
+        if attention_mask is not None:
+            # expand our mask's singleton query_tokens dimension:
+            #   [batch*heads,            1, key_tokens] ->
+            #   [batch*heads, query_tokens, key_tokens]
+            # so that it can be added as a bias onto the attention scores that xformers computes:
+            #   [batch*heads, query_tokens, key_tokens]
+            # we do this explicitly because xformers doesn't broadcast the singleton dimension for us.
+            _, query_tokens, _ = hidden_states.shape
+            attention_mask = attention_mask.expand(-1, query_tokens, -1)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        assert num_tasks == 2  # only support two tasks now
+
+        key_0, key_1 = torch.chunk(key, dim=0, chunks=2)  # keys shape (b t) d c
+        value_0, value_1 = torch.chunk(value, dim=0, chunks=2)
+
+        # key = torch.cat([key_1, key_0], dim=0)
+        # value = torch.cat([value_1, value_0], dim=0)
+
+        key = torch.cat([key_0, key_1], dim=1)  # (b t) 2d c
+        value = torch.cat([value_0, value_1], dim=1)  # (b t) 2d c
+        key = torch.cat([key]*2, dim=0)   # (2 b t) 2d c
+        value = torch.cat([value]*2, dim=0)  # (2 b t) 2d c
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        hidden_states = xformers.ops.memory_efficient_attention(query, key, value, attn_bias=attention_mask)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+        
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm_in = nn.LayerNorm(dim)
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = LoRACompatibleLinear if not USE_PEFT_BACKEND else nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out, bias=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, scale: float = 1.0) -> torch.Tensor:
+        compatible_cls = (GEGLU,) if USE_PEFT_BACKEND else (GEGLU, LoRACompatibleLinear)
+        for module in self.net:
+            if isinstance(module, compatible_cls):
+                hidden_states = module(hidden_states, scale)
+            else:
+                hidden_states = module(hidden_states)
+        return hidden_states

models/depth_normal_pipeline.py

@@ -0,0 +1,361 @@
+# A reimplemented version in public environments by Xiao Fu and Mu Hu
+
+from typing import Any, Dict, Union
+
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+import numpy as np
+from tqdm.auto import tqdm
+from PIL import Image
+from diffusers import (
+    DiffusionPipeline,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from models.unet_2d_condition import UNet2DConditionModel
+from diffusers.utils import BaseOutput
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from utils.image_util import resize_max_res,chw2hwc,colorize_depth_maps
+from utils.colormap import kitti_colormap
+from utils.depth_ensemble import ensemble_depths
+from utils.batch_size import find_batch_size
+import cv2
+
+class DepthNormalPipelineOutput(BaseOutput):
+    """
+    Output class for Marigold monocular depth prediction pipeline.
+
+    Args:
+        depth_np (`np.ndarray`):
+            Predicted depth map, with depth values in the range of [0, 1].
+        depth_colored (`PIL.Image.Image`):
+            Colorized depth map, with the shape of [3, H, W] and values in [0, 1].
+        normal_np (`np.ndarray`):
+            Predicted normal map, with depth values in the range of [0, 1].
+        normal_colored (`PIL.Image.Image`):
+            Colorized normal map, with the shape of [3, H, W] and values in [0, 1].
+        uncertainty (`None` or `np.ndarray`):
+            Uncalibrated uncertainty(MAD, median absolute deviation) coming from ensembling.
+    """
+    depth_np: np.ndarray
+    depth_colored: Image.Image
+    normal_np: np.ndarray
+    normal_colored: Image.Image
+    uncertainty: Union[None, np.ndarray]
+
+class DepthNormalEstimationPipeline(DiffusionPipeline):
+    # two hyper-parameters
+    latent_scale_factor = 0.18215
+
+    def __init__(self,
+                 unet:UNet2DConditionModel,
+                 vae:AutoencoderKL,
+                 scheduler:DDIMScheduler,
+                 text_encoder:CLIPTextModel,
+                 tokenizer:CLIPTokenizer,
+                 ):
+        super().__init__()
+            
+        self.register_modules(
+            unet=unet,
+            vae=vae,
+            scheduler=scheduler,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+        )
+        self.empty_text_embed = None  
+    
+    @torch.no_grad()
+    def __call__(self,
+                 input_image:Image,
+                 denoising_steps: int = 10,
+                 ensemble_size: int = 10,
+                 processing_res: int = 768,
+                 match_input_res:bool =True,
+                 batch_size:int = 0,
+                 domain: str = "indoor",
+                 color_map: str="Spectral",
+                 show_progress_bar:bool = True,
+                 ensemble_kwargs: Dict = None,
+                 ) -> DepthNormalPipelineOutput:
+        
+        # inherit from thea Diffusion Pipeline
+        device = self.device
+        input_size = input_image.size
+        
+        # adjust the input resolution.
+        if not match_input_res:
+            assert (
+                processing_res is not None                
+            )," Value Error: `resize_output_back` is only valid with "
+        
+        assert processing_res >=0
+        assert denoising_steps >=1
+        assert ensemble_size >=1
+
+        # --------------- Image Processing ------------------------
+        # Resize image
+        if processing_res >0:
+            input_image = resize_max_res(
+                input_image, max_edge_resolution=processing_res
+            )
+        
+        # Convert the image to RGB, to 1. reomve the alpha channel.
+        input_image = input_image.convert("RGB")
+        image = np.array(input_image)
+
+        # Normalize RGB Values.
+        rgb = np.transpose(image,(2,0,1))
+        rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
+        rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
+        rgb_norm = rgb_norm.to(device)
+
+        assert rgb_norm.min() >= -1.0 and rgb_norm.max() <= 1.0
+        
+        # ----------------- predicting depth -----------------
+        duplicated_rgb = torch.stack([rgb_norm] * ensemble_size)
+        single_rgb_dataset = TensorDataset(duplicated_rgb)
+        
+        # find the batch size
+        if batch_size>0:
+            _bs = batch_size
+        else:
+            _bs = 1
+
+        single_rgb_loader = DataLoader(single_rgb_dataset, batch_size=_bs, shuffle=False)
+        
+        # predicted the depth
+        depth_pred_ls = []
+        normal_pred_ls = []
+        
+        if show_progress_bar:
+            iterable_bar = tqdm(
+                single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
+            )
+        else:
+            iterable_bar = single_rgb_loader
+        
+        for batch in iterable_bar:
+            (batched_image, )= batch  # here the image is still around 0-1
+
+            depth_pred_raw, normal_pred_raw = self.single_infer(
+                input_rgb=batched_image,
+                num_inference_steps=denoising_steps,
+                domain=domain,
+                show_pbar=show_progress_bar,
+            )
+            depth_pred_ls.append(depth_pred_raw.detach().clone())
+            normal_pred_ls.append(normal_pred_raw.detach().clone())
+        
+        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze() #(10,224,768)
+        normal_preds = torch.concat(normal_pred_ls, axis=0).squeeze()
+        torch.cuda.empty_cache()  # clear vram cache for ensembling
+
+        # ----------------- Test-time ensembling -----------------
+        if ensemble_size > 1:
+            depth_pred, pred_uncert = ensemble_depths(
+                depth_preds, **(ensemble_kwargs or {})
+            )
+            normal_pred = normal_preds[0]
+        else:
+            depth_pred = depth_preds
+            normal_pred = normal_preds
+            pred_uncert = None
+
+        # ----------------- Post processing -----------------
+        # Scale prediction to [0, 1]
+        min_d = torch.min(depth_pred)
+        max_d = torch.max(depth_pred)
+        depth_pred = (depth_pred - min_d) / (max_d - min_d)
+        
+        # Convert to numpy
+        depth_pred = depth_pred.cpu().numpy().astype(np.float32)
+        normal_pred = normal_pred.cpu().numpy().astype(np.float32)
+
+        # Resize back to original resolution
+        if match_input_res:
+            pred_img = Image.fromarray(depth_pred)
+            pred_img = pred_img.resize(input_size)
+            depth_pred = np.asarray(pred_img)
+            normal_pred = cv2.resize(chw2hwc(normal_pred), input_size, interpolation = cv2.INTER_NEAREST)
+
+        # Clip output range: current size is the original size
+        depth_pred = depth_pred.clip(0, 1)
+        normal_pred = normal_pred.clip(-1, 1)
+        
+        # Colorize
+        depth_colored = colorize_depth_maps(
+            depth_pred, 0, 1, cmap=color_map
+        ).squeeze()  # [3, H, W], value in (0, 1)
+        depth_colored = (depth_colored * 255).astype(np.uint8)
+        depth_colored_hwc = chw2hwc(depth_colored)
+        depth_colored_img = Image.fromarray(depth_colored_hwc)
+
+        normal_colored = ((normal_pred + 1)/2 * 255).astype(np.uint8)
+        normal_colored_img = Image.fromarray(normal_colored)
+        
+        return DepthNormalPipelineOutput(
+            depth_np = depth_pred,
+            depth_colored = depth_colored_img,
+            normal_np = normal_pred,
+            normal_colored = normal_colored_img,
+            uncertainty=pred_uncert,
+        )
+    
+    def __encode_empty_text(self):
+        """
+        Encode text embedding for empty prompt
+        """
+        prompt = ""
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="do_not_pad",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids.to(self.text_encoder.device) #[1,2]
+        # print(text_input_ids.shape)
+        self.empty_text_embed = self.text_encoder(text_input_ids)[0].to(self.dtype) #[1,2,1024]
+
+        
+    @torch.no_grad()
+    def single_infer(self,input_rgb:torch.Tensor,
+                     num_inference_steps:int,
+                     domain:str,
+                     show_pbar:bool,):
+
+        device = input_rgb.device
+
+        # Set timesteps: inherit from the diffuison pipeline
+        self.scheduler.set_timesteps(num_inference_steps, device=device) # here the numbers of the steps is only 10.
+        timesteps = self.scheduler.timesteps  # [T]
+        
+        # encode image
+        rgb_latent = self.encode_RGB(input_rgb)
+        
+        # Initial depth map (Guassian noise)
+        geo_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype).repeat(2,1,1,1)
+        rgb_latent = rgb_latent.repeat(2,1,1,1)
+
+        # Batched empty text embedding
+        if self.empty_text_embed is None:
+            self.__encode_empty_text()
+        
+        batch_empty_text_embed = self.empty_text_embed.repeat(
+            (rgb_latent.shape[0], 1, 1)
+        )  # [B, 2, 1024]
+
+        # hybrid hierarchical switcher 
+        geo_class = torch.tensor([[0., 1.], [1, 0]], device=device, dtype=self.dtype)
+        geo_embedding = torch.cat([torch.sin(geo_class), torch.cos(geo_class)], dim=-1)
+        
+        if domain == "indoor":
+            domain_class = torch.tensor([[1., 0., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "outdoor":
+            domain_class = torch.tensor([[0., 1., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "object":
+            domain_class = torch.tensor([[0., 0., 1]], device=device, dtype=self.dtype).repeat(2,1)
+        domain_embedding = torch.cat([torch.sin(domain_class), torch.cos(domain_class)], dim=-1)
+
+        class_embedding = torch.cat((geo_embedding, domain_embedding), dim=-1)
+
+        # Denoising loop
+        if show_pbar:
+            iterable = tqdm(
+                enumerate(timesteps),
+                total=len(timesteps),
+                leave=False,
+                desc=" " * 4 + "Diffusion denoising",
+            )
+        else:
+            iterable = enumerate(timesteps)
+        
+        for i, t in iterable:
+            unet_input = torch.cat([rgb_latent, geo_latent], dim=1)
+
+            # predict the noise residual
+            noise_pred = self.unet(
+                unet_input, t.repeat(2), encoder_hidden_states=batch_empty_text_embed, class_labels=class_embedding
+            ).sample  # [B, 4, h, w]
+
+            # compute the previous noisy sample x_t -> x_t-1
+            geo_latent = self.scheduler.step(noise_pred, t, geo_latent).prev_sample
+
+        geo_latent = geo_latent
+        torch.cuda.empty_cache()
+
+        depth = self.decode_depth(geo_latent[0][None])
+        depth = torch.clip(depth, -1.0, 1.0)
+        depth = (depth + 1.0) / 2.0
+        
+        normal = self.decode_normal(geo_latent[1][None])
+        normal /= (torch.norm(normal, p=2, dim=1, keepdim=True)+1e-5)
+
+        return depth, normal
+        
+    
+    def encode_RGB(self, rgb_in: torch.Tensor) -> torch.Tensor:
+        """
+        Encode RGB image into latent.
+
+        Args:
+            rgb_in (`torch.Tensor`):
+                Input RGB image to be encoded.
+
+        Returns:
+            `torch.Tensor`: Image latent.
+        """
+
+        # encode
+        h = self.vae.encoder(rgb_in)
+
+        moments = self.vae.quant_conv(h)
+        mean, logvar = torch.chunk(moments, 2, dim=1)
+        # scale latent
+        rgb_latent = mean * self.latent_scale_factor
+        
+        return rgb_latent
+    
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode depth latent into depth map.
+
+        Args:
+            depth_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded depth map.
+        """
+
+        # scale latent
+        depth_latent = depth_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(depth_latent)
+        stacked = self.vae.decoder(z)
+        # mean of output channels
+        depth_mean = stacked.mean(dim=1, keepdim=True)
+        return depth_mean
+
+    def decode_normal(self, normal_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode normal latent into normal map.
+
+        Args:
+            normal_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded normal map.
+        """
+
+        # scale latent
+        normal_latent = normal_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(normal_latent)
+        normal = self.vae.decoder(z)
+        return normal
+
+

models/depth_normal_pipeline_clip.py

@@ -0,0 +1,368 @@
+# A reimplemented version in public environments by Xiao Fu and Mu Hu
+
+from typing import Any, Dict, Union
+
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+import numpy as np
+from tqdm.auto import tqdm
+from PIL import Image
+from diffusers import (
+    DiffusionPipeline,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from models.unet_2d_condition import UNet2DConditionModel
+from diffusers.utils import BaseOutput
+from transformers import CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+import torchvision.transforms.functional as TF
+from torchvision.transforms import InterpolationMode
+
+from utils.image_util import resize_max_res,chw2hwc,colorize_depth_maps
+from utils.colormap import kitti_colormap
+from utils.depth_ensemble import ensemble_depths
+from utils.normal_ensemble import ensemble_normals
+from utils.batch_size import find_batch_size
+import cv2
+
+class DepthNormalPipelineOutput(BaseOutput):
+    """
+    Output class for Marigold monocular depth prediction pipeline.
+
+    Args:
+        depth_np (`np.ndarray`):
+            Predicted depth map, with depth values in the range of [0, 1].
+        depth_colored (`PIL.Image.Image`):
+            Colorized depth map, with the shape of [3, H, W] and values in [0, 1].
+        normal_np (`np.ndarray`):
+            Predicted normal map, with depth values in the range of [0, 1].
+        normal_colored (`PIL.Image.Image`):
+            Colorized normal map, with the shape of [3, H, W] and values in [0, 1].
+        uncertainty (`None` or `np.ndarray`):
+            Uncalibrated uncertainty(MAD, median absolute deviation) coming from ensembling.
+    """
+    depth_np: np.ndarray
+    depth_colored: Image.Image
+    normal_np: np.ndarray
+    normal_colored: Image.Image
+    uncertainty: Union[None, np.ndarray]
+
+class DepthNormalEstimationPipeline(DiffusionPipeline):
+    # two hyper-parameters
+    latent_scale_factor = 0.18215
+
+    def __init__(self,
+                 unet:UNet2DConditionModel,
+                 vae:AutoencoderKL,
+                 scheduler:DDIMScheduler,
+                 image_encoder:CLIPVisionModelWithProjection,
+                 feature_extractor:CLIPImageProcessor,
+                 ):
+        super().__init__()
+            
+        self.register_modules(
+            unet=unet,
+            vae=vae,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.img_embed = None  
+
+    @torch.no_grad()
+    def __call__(self,
+                 input_image:Image,
+                 denosing_steps: int = 10,
+                 ensemble_size: int = 10,
+                 processing_res: int = 768,
+                 match_input_res:bool =True,
+                 batch_size:int = 0,
+                 domain: str = "indoor",
+                 color_map: str="Spectral",
+                 show_progress_bar:bool = True,
+                 ensemble_kwargs: Dict = None,
+                 ) -> DepthNormalPipelineOutput:
+        
+        # inherit from thea Diffusion Pipeline
+        device = self.device
+        input_size = input_image.size
+        
+        # adjust the input resolution.
+        if not match_input_res:
+            assert (
+                processing_res is not None                
+            )," Value Error: `resize_output_back` is only valid with "
+        
+        assert processing_res >=0
+        assert denosing_steps >=1
+        assert ensemble_size >=1
+
+        # --------------- Image Processing ------------------------
+        # Resize image
+        if processing_res >0:
+            input_image = resize_max_res(
+                input_image, max_edge_resolution=processing_res
+            )
+        
+        # Convert the image to RGB, to 1. reomve the alpha channel.
+        input_image = input_image.convert("RGB")
+        image = np.array(input_image)
+
+        # Normalize RGB Values.
+        rgb = np.transpose(image,(2,0,1))
+        rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
+        rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
+        rgb_norm = rgb_norm.to(device)
+
+        assert rgb_norm.min() >= -1.0 and rgb_norm.max() <= 1.0
+        
+        # ----------------- predicting depth -----------------
+        duplicated_rgb = torch.stack([rgb_norm] * ensemble_size)
+        single_rgb_dataset = TensorDataset(duplicated_rgb)
+        
+        # find the batch size
+        if batch_size>0:
+            _bs = batch_size
+        else:
+            _bs = 1
+
+        single_rgb_loader = DataLoader(single_rgb_dataset, batch_size=_bs, shuffle=False)
+        
+        # predicted the depth
+        depth_pred_ls = []
+        normal_pred_ls = []
+        
+        if show_progress_bar:
+            iterable_bar = tqdm(
+                single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
+            )
+        else:
+            iterable_bar = single_rgb_loader
+        
+        for batch in iterable_bar:
+            (batched_image, )= batch  # here the image is still around 0-1
+
+            depth_pred_raw, normal_pred_raw = self.single_infer(
+                input_rgb=batched_image,
+                num_inference_steps=denosing_steps,
+                domain=domain,
+                show_pbar=show_progress_bar,
+            )
+            depth_pred_ls.append(depth_pred_raw.detach().clone())
+            normal_pred_ls.append(normal_pred_raw.detach().clone())
+        
+        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze() #(10,224,768)
+        normal_preds = torch.concat(normal_pred_ls, axis=0).squeeze()
+        torch.cuda.empty_cache()  # clear vram cache for ensembling
+
+        # ----------------- Test-time ensembling -----------------
+        if ensemble_size > 1:
+            depth_pred, pred_uncert = ensemble_depths(
+                depth_preds, **(ensemble_kwargs or {})
+            )
+            normal_pred = ensemble_normals(normal_preds)
+        else:
+            depth_pred = depth_preds
+            normal_pred = normal_preds
+            pred_uncert = None
+
+        # ----------------- Post processing -----------------
+        # Scale prediction to [0, 1]
+        min_d = torch.min(depth_pred)
+        max_d = torch.max(depth_pred)
+        depth_pred = (depth_pred - min_d) / (max_d - min_d)
+        
+        # Convert to numpy
+        depth_pred = depth_pred.cpu().numpy().astype(np.float32)
+        normal_pred = normal_pred.cpu().numpy().astype(np.float32)
+
+        # Resize back to original resolution
+        if match_input_res:
+            pred_img = Image.fromarray(depth_pred)
+            pred_img = pred_img.resize(input_size)
+            depth_pred = np.asarray(pred_img)
+            normal_pred = cv2.resize(chw2hwc(normal_pred), input_size, interpolation = cv2.INTER_NEAREST)
+
+        # Clip output range: current size is the original size
+        depth_pred = depth_pred.clip(0, 1)
+        normal_pred = normal_pred.clip(-1, 1)
+    
+        # Colorize
+        depth_colored = colorize_depth_maps(
+            depth_pred, 0, 1, cmap=color_map
+        ).squeeze()  # [3, H, W], value in (0, 1)
+        depth_colored = (depth_colored * 255).astype(np.uint8)
+        depth_colored_hwc = chw2hwc(depth_colored)
+        depth_colored_img = Image.fromarray(depth_colored_hwc)
+
+        normal_colored = ((normal_pred + 1)/2 * 255).astype(np.uint8)
+        normal_colored_img = Image.fromarray(normal_colored)
+        
+        return DepthNormalPipelineOutput(
+            depth_np = depth_pred,
+            depth_colored = depth_colored_img,
+            normal_np = normal_pred,
+            normal_colored = normal_colored_img,
+            uncertainty=pred_uncert,
+        )
+    
+    def __encode_img_embed(self, rgb):
+        """
+        Encode clip embeddings for img
+        """
+        clip_image_mean = torch.as_tensor(self.feature_extractor.image_mean)[:,None,None].to(device=self.device, dtype=self.dtype)
+        clip_image_std = torch.as_tensor(self.feature_extractor.image_std)[:,None,None].to(device=self.device, dtype=self.dtype)
+
+        img_in_proc = TF.resize((rgb +1)/2, 
+            (self.feature_extractor.crop_size['height'], self.feature_extractor.crop_size['width']), 
+            interpolation=InterpolationMode.BICUBIC, 
+            antialias=True
+        )
+        # do the normalization in float32 to preserve precision
+        img_in_proc = ((img_in_proc.float() - clip_image_mean) / clip_image_std).to(self.dtype)        
+        img_embed = self.image_encoder(img_in_proc).image_embeds.unsqueeze(1).to(self.dtype)
+
+        self.img_embed = img_embed
+
+        
+    @torch.no_grad()
+    def single_infer(self,input_rgb:torch.Tensor,
+                     num_inference_steps:int,
+                     domain:str,
+                     show_pbar:bool,):
+
+        device = input_rgb.device
+
+        # Set timesteps: inherit from the diffuison pipeline
+        self.scheduler.set_timesteps(num_inference_steps, device=device) # here the numbers of the steps is only 10.
+        timesteps = self.scheduler.timesteps  # [T]
+        
+        # encode image
+        rgb_latent = self.encode_RGB(input_rgb)
+        
+        # Initial depth map (Guassian noise)
+        geo_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype).repeat(2,1,1,1)
+        rgb_latent = rgb_latent.repeat(2,1,1,1)
+
+        # Batched img embedding
+        if self.img_embed is None:
+            self.__encode_img_embed(input_rgb)
+        
+        batch_img_embed = self.img_embed.repeat(
+            (rgb_latent.shape[0], 1, 1)
+        )  # [B, 1, 768]
+
+        # hybrid hierarchical switcher 
+        geo_class = torch.tensor([[0., 1.], [1, 0]], device=device, dtype=self.dtype)
+        geo_embedding = torch.cat([torch.sin(geo_class), torch.cos(geo_class)], dim=-1)
+        
+        if domain == "indoor":
+            domain_class = torch.tensor([[1., 0., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "outdoor":
+            domain_class = torch.tensor([[0., 1., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "object":
+            domain_class = torch.tensor([[0., 0., 1]], device=device, dtype=self.dtype).repeat(2,1)
+        domain_embedding = torch.cat([torch.sin(domain_class), torch.cos(domain_class)], dim=-1)
+
+        class_embedding = torch.cat((geo_embedding, domain_embedding), dim=-1)
+
+        # Denoising loop
+        if show_pbar:
+            iterable = tqdm(
+                enumerate(timesteps),
+                total=len(timesteps),
+                leave=False,
+                desc=" " * 4 + "Diffusion denoising",
+            )
+        else:
+            iterable = enumerate(timesteps)
+        
+        for i, t in iterable:
+            unet_input = torch.cat([rgb_latent, geo_latent], dim=1)
+
+            # predict the noise residual
+            noise_pred = self.unet(
+                unet_input, t.repeat(2), encoder_hidden_states=batch_img_embed, class_labels=class_embedding
+            ).sample  # [B, 4, h, w]
+
+            # compute the previous noisy sample x_t -> x_t-1
+            geo_latent = self.scheduler.step(noise_pred, t, geo_latent).prev_sample
+
+        geo_latent = geo_latent
+        torch.cuda.empty_cache()
+
+        depth = self.decode_depth(geo_latent[0][None])
+        depth = torch.clip(depth, -1.0, 1.0)
+        depth = (depth + 1.0) / 2.0
+        
+        normal = self.decode_normal(geo_latent[1][None])
+        normal /= (torch.norm(normal, p=2, dim=1, keepdim=True)+1e-5)
+        normal *= -1.
+
+        return depth, normal
+        
+    
+    def encode_RGB(self, rgb_in: torch.Tensor) -> torch.Tensor:
+        """
+        Encode RGB image into latent.
+
+        Args:
+            rgb_in (`torch.Tensor`):
+                Input RGB image to be encoded.
+
+        Returns:
+            `torch.Tensor`: Image latent.
+        """
+
+        # encode
+        h = self.vae.encoder(rgb_in)
+
+        moments = self.vae.quant_conv(h)
+        mean, logvar = torch.chunk(moments, 2, dim=1)
+        # scale latent
+        rgb_latent = mean * self.latent_scale_factor
+        
+        return rgb_latent
+    
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode depth latent into depth map.
+
+        Args:
+            depth_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded depth map.
+        """
+
+        # scale latent
+        depth_latent = depth_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(depth_latent)
+        stacked = self.vae.decoder(z)
+        # mean of output channels
+        depth_mean = stacked.mean(dim=1, keepdim=True)
+        return depth_mean
+
+    def decode_normal(self, normal_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode normal latent into normal map.
+
+        Args:
+            normal_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded normal map.
+        """
+
+        # scale latent
+        normal_latent = normal_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(normal_latent)
+        normal = self.vae.decoder(z)
+        return normal
+
+

models/depth_normal_pipeline_clip_cfg.py

@@ -0,0 +1,371 @@
+# A reimplemented version in public environments by Xiao Fu and Mu Hu
+
+from typing import Any, Dict, Union
+
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+import numpy as np
+from tqdm.auto import tqdm
+from PIL import Image
+from diffusers import (
+    DiffusionPipeline,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from models.unet_2d_condition import UNet2DConditionModel
+from diffusers.utils import BaseOutput
+from transformers import CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+import torchvision.transforms.functional as TF
+from torchvision.transforms import InterpolationMode
+
+from utils.image_util import resize_max_res,chw2hwc,colorize_depth_maps
+from utils.colormap import kitti_colormap
+from utils.depth_ensemble import ensemble_depths
+from utils.normal_ensemble import ensemble_normals
+from utils.batch_size import find_batch_size
+import cv2
+
+class DepthNormalPipelineOutput(BaseOutput):
+    """
+    Output class for Marigold monocular depth prediction pipeline.
+
+    Args:
+        depth_np (`np.ndarray`):
+            Predicted depth map, with depth values in the range of [0, 1].
+        depth_colored (`PIL.Image.Image`):
+            Colorized depth map, with the shape of [3, H, W] and values in [0, 1].
+        normal_np (`np.ndarray`):
+            Predicted normal map, with depth values in the range of [0, 1].
+        normal_colored (`PIL.Image.Image`):
+            Colorized normal map, with the shape of [3, H, W] and values in [0, 1].
+        uncertainty (`None` or `np.ndarray`):
+            Uncalibrated uncertainty(MAD, median absolute deviation) coming from ensembling.
+    """
+    depth_np: np.ndarray
+    depth_colored: Image.Image
+    normal_np: np.ndarray
+    normal_colored: Image.Image
+    uncertainty: Union[None, np.ndarray]
+
+class DepthNormalEstimationPipeline(DiffusionPipeline):
+    # two hyper-parameters
+    latent_scale_factor = 0.18215
+
+    def __init__(self,
+                 unet:UNet2DConditionModel,
+                 vae:AutoencoderKL,
+                 scheduler:DDIMScheduler,
+                 image_encoder:CLIPVisionModelWithProjection,
+                 feature_extractor:CLIPImageProcessor,
+                 ):
+        super().__init__()
+            
+        self.register_modules(
+            unet=unet,
+            vae=vae,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.img_embed = None  
+
+    @torch.no_grad()
+    def __call__(self,
+                 input_image:Image,
+                 denosing_steps: int = 10,
+                 ensemble_size: int = 10,
+                 processing_res: int = 768,
+                 match_input_res:bool =True,
+                 batch_size:int = 0,
+                 domain: str = "indoor",
+                 color_map: str="Spectral",
+                 show_progress_bar:bool = True,
+                 ensemble_kwargs: Dict = None,
+                 ) -> DepthNormalPipelineOutput:
+        
+        # inherit from thea Diffusion Pipeline
+        device = self.device
+        input_size = input_image.size
+        
+        # adjust the input resolution.
+        if not match_input_res:
+            assert (
+                processing_res is not None                
+            )," Value Error: `resize_output_back` is only valid with "
+        
+        assert processing_res >=0
+        assert denosing_steps >=1
+        assert ensemble_size >=1
+
+        # --------------- Image Processing ------------------------
+        # Resize image
+        if processing_res >0:
+            input_image = resize_max_res(
+                input_image, max_edge_resolution=processing_res
+            )
+        
+        # Convert the image to RGB, to 1. reomve the alpha channel.
+        input_image = input_image.convert("RGB")
+        image = np.array(input_image)
+
+        # Normalize RGB Values.
+        rgb = np.transpose(image,(2,0,1))
+        rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
+        rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
+        rgb_norm = rgb_norm.to(device)
+
+        assert rgb_norm.min() >= -1.0 and rgb_norm.max() <= 1.0
+        
+        # ----------------- predicting depth -----------------
+        duplicated_rgb = torch.stack([rgb_norm] * ensemble_size)
+        single_rgb_dataset = TensorDataset(duplicated_rgb)
+        
+        # find the batch size
+        if batch_size>0:
+            _bs = batch_size
+        else:
+            _bs = 1
+
+        single_rgb_loader = DataLoader(single_rgb_dataset, batch_size=_bs, shuffle=False)
+        
+        # predicted the depth
+        depth_pred_ls = []
+        normal_pred_ls = []
+        
+        if show_progress_bar:
+            iterable_bar = tqdm(
+                single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
+            )
+        else:
+            iterable_bar = single_rgb_loader
+        
+        for batch in iterable_bar:
+            (batched_image, )= batch  # here the image is still around 0-1
+
+            depth_pred_raw, normal_pred_raw = self.single_infer(
+                input_rgb=batched_image,
+                num_inference_steps=denosing_steps,
+                domain=domain,
+                show_pbar=show_progress_bar,
+            )
+            depth_pred_ls.append(depth_pred_raw.detach().clone())
+            normal_pred_ls.append(normal_pred_raw.detach().clone())
+        
+        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze()
+        normal_preds = torch.concat(normal_pred_ls, axis=0).squeeze()
+        torch.cuda.empty_cache()  # clear vram cache for ensembling
+
+        # ----------------- Test-time ensembling -----------------
+        if ensemble_size > 1:
+            depth_pred, pred_uncert = ensemble_depths(
+                depth_preds, **(ensemble_kwargs or {})
+            )
+            normal_pred = ensemble_normals(normal_preds)
+        else:
+            depth_pred = depth_preds
+            normal_pred = normal_preds
+            pred_uncert = None
+
+        # ----------------- Post processing -----------------
+        # Scale prediction to [0, 1]
+        min_d = torch.min(depth_pred)
+        max_d = torch.max(depth_pred)
+        depth_pred = (depth_pred - min_d) / (max_d - min_d)
+        
+        # Convert to numpy
+        depth_pred = depth_pred.cpu().numpy().astype(np.float32)
+        normal_pred = normal_pred.cpu().numpy().astype(np.float32)
+
+        # Resize back to original resolution
+        if match_input_res:
+            pred_img = Image.fromarray(depth_pred)
+            pred_img = pred_img.resize(input_size)
+            depth_pred = np.asarray(pred_img)
+            normal_pred = cv2.resize(chw2hwc(normal_pred), input_size, interpolation = cv2.INTER_NEAREST)
+
+        # Clip output range: current size is the original size
+        depth_pred = depth_pred.clip(0, 1)
+        normal_pred = normal_pred.clip(-1, 1)
+    
+        # Colorize
+        depth_colored = colorize_depth_maps(
+            depth_pred, 0, 1, cmap=color_map
+        ).squeeze()  # [3, H, W], value in (0, 1)
+        depth_colored = (depth_colored * 255).astype(np.uint8)
+        depth_colored_hwc = chw2hwc(depth_colored)
+        depth_colored_img = Image.fromarray(depth_colored_hwc)
+
+        normal_colored = ((normal_pred + 1)/2 * 255).astype(np.uint8)
+        normal_colored_img = Image.fromarray(normal_colored)
+        
+        return DepthNormalPipelineOutput(
+            depth_np = depth_pred,
+            depth_colored = depth_colored_img,
+            normal_np = normal_pred,
+            normal_colored = normal_colored_img,
+            uncertainty=pred_uncert,
+        )
+    
+    def __encode_img_embed(self, rgb):
+        """
+        Encode clip embeddings for img
+        """
+        clip_image_mean = torch.as_tensor(self.feature_extractor.image_mean)[:,None,None].to(device=self.device, dtype=self.dtype)
+        clip_image_std = torch.as_tensor(self.feature_extractor.image_std)[:,None,None].to(device=self.device, dtype=self.dtype)
+
+        img_in_proc = TF.resize((rgb +1)/2, 
+            (self.feature_extractor.crop_size['height'], self.feature_extractor.crop_size['width']), 
+            interpolation=InterpolationMode.BICUBIC, 
+            antialias=True
+        )
+        # do the normalization in float32 to preserve precision
+        img_in_proc = ((img_in_proc.float() - clip_image_mean) / clip_image_std).to(self.dtype)        
+        img_embed = self.image_encoder(img_in_proc).image_embeds.unsqueeze(1).to(self.dtype)
+
+        self.img_embed = img_embed
+
+        
+    @torch.no_grad()
+    def single_infer(self,input_rgb:torch.Tensor,
+                     num_inference_steps:int,
+                     domain:str,
+                     show_pbar:bool,):
+
+        device = input_rgb.device
+
+        # Set timesteps: inherit from the diffuison pipeline
+        self.scheduler.set_timesteps(num_inference_steps, device=device) # here the numbers of the steps is only 10.
+        timesteps = self.scheduler.timesteps  # [T]
+        
+        # encode image
+        rgb_latent = self.encode_RGB(input_rgb)
+        
+        # Initial depth map (Guassian noise)
+        geo_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype).repeat(2,1,1,1)
+        rgb_latent = rgb_latent.repeat(2,1,1,1)
+
+        # Batched img embedding
+        if self.img_embed is None:
+            self.__encode_img_embed(input_rgb)
+        
+        batch_img_embed = self.img_embed.repeat(
+            (rgb_latent.shape[0], 1, 1)
+        )  # [B, 1, 768]
+
+        batch_img_embed = torch.cat((torch.zeros_like(batch_img_embed), batch_img_embed), dim=0)
+        rgb_latent = torch.cat((torch.zeros_like(rgb_latent), rgb_latent), dim=0)
+
+        # hybrid switcher 
+        geo_class = torch.tensor([[0., 1.], [1, 0]], device=device, dtype=self.dtype)
+        geo_embedding = torch.cat([torch.sin(geo_class), torch.cos(geo_class)], dim=-1)
+        
+        if domain == "indoor":
+            domain_class = torch.tensor([[1., 0., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "outdoor":
+            domain_class = torch.tensor([[0., 1., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "object":
+            domain_class = torch.tensor([[0., 0., 1]], device=device, dtype=self.dtype).repeat(2,1)
+        domain_embedding = torch.cat([torch.sin(domain_class), torch.cos(domain_class)], dim=-1)
+
+        class_embedding = torch.cat((geo_embedding, domain_embedding), dim=-1)
+
+        # Denoising loop
+        if show_pbar:
+            iterable = tqdm(
+                enumerate(timesteps),
+                total=len(timesteps),
+                leave=False,
+                desc=" " * 4 + "Diffusion denoising",
+            )
+        else:
+            iterable = enumerate(timesteps)
+
+        for i, t in iterable:
+            unet_input = torch.cat((rgb_latent, geo_latent.repeat(2,1,1,1)), dim=1)
+            # predict the noise residual
+            noise_pred = self.unet(unet_input, t.repeat(4), encoder_hidden_states=batch_img_embed, class_labels=class_embedding.repeat(2,1)).sample  
+            noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
+            guidance_scale = 3.
+            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
+        
+            # compute the previous noisy sample x_t -> x_t-1
+            geo_latent = self.scheduler.step(noise_pred, t, geo_latent).prev_sample
+
+        geo_latent = geo_latent
+        torch.cuda.empty_cache()
+
+        depth = self.decode_depth(geo_latent[0][None])
+        depth = torch.clip(depth, -1.0, 1.0)
+        depth = (depth + 1.0) / 2.0
+        
+        normal = self.decode_normal(geo_latent[1][None])
+        normal /= (torch.norm(normal, p=2, dim=1, keepdim=True)+1e-5)
+        normal *= -1.
+
+        return depth, normal
+        
+    
+    def encode_RGB(self, rgb_in: torch.Tensor) -> torch.Tensor:
+        """
+        Encode RGB image into latent.
+
+        Args:
+            rgb_in (`torch.Tensor`):
+                Input RGB image to be encoded.
+
+        Returns:
+            `torch.Tensor`: Image latent.
+        """
+
+        # encode
+        h = self.vae.encoder(rgb_in)
+
+        moments = self.vae.quant_conv(h)
+        mean, logvar = torch.chunk(moments, 2, dim=1)
+        # scale latent
+        rgb_latent = mean * self.latent_scale_factor
+        
+        return rgb_latent
+    
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode depth latent into depth map.
+
+        Args:
+            depth_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded depth map.
+        """
+
+        # scale latent
+        depth_latent = depth_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(depth_latent)
+        stacked = self.vae.decoder(z)
+        # mean of output channels
+        depth_mean = stacked.mean(dim=1, keepdim=True)
+        return depth_mean
+
+    def decode_normal(self, normal_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode normal latent into normal map.
+
+        Args:
+            normal_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded normal map.
+        """
+
+        # scale latent
+        normal_latent = normal_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(normal_latent)
+        normal = self.vae.decoder(z)
+        return normal
+
+

models/depth_normal_pipeline_clip_cfg_1.py

@@ -0,0 +1,371 @@
+# A reimplemented version in public environments by Xiao Fu and Mu Hu
+
+from typing import Any, Dict, Union
+
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+import numpy as np
+from tqdm.auto import tqdm
+from PIL import Image
+from diffusers import (
+    DiffusionPipeline,
+    DDIMScheduler,
+    AutoencoderKL,
+)
+from models.unet_2d_condition import UNet2DConditionModel
+from diffusers.utils import BaseOutput
+from transformers import CLIPTextModel, CLIPTokenizer
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+import torchvision.transforms.functional as TF
+from torchvision.transforms import InterpolationMode
+
+from utils.image_util import resize_max_res,chw2hwc,colorize_depth_maps
+from utils.colormap import kitti_colormap
+from utils.depth_ensemble import ensemble_depths
+from utils.normal_ensemble import ensemble_normals
+from utils.batch_size import find_batch_size
+import cv2
+
+class DepthNormalPipelineOutput(BaseOutput):
+    """
+    Output class for Marigold monocular depth prediction pipeline.
+
+    Args:
+        depth_np (`np.ndarray`):
+            Predicted depth map, with depth values in the range of [0, 1].
+        depth_colored (`PIL.Image.Image`):
+            Colorized depth map, with the shape of [3, H, W] and values in [0, 1].
+        normal_np (`np.ndarray`):
+            Predicted normal map, with depth values in the range of [0, 1].
+        normal_colored (`PIL.Image.Image`):
+            Colorized normal map, with the shape of [3, H, W] and values in [0, 1].
+        uncertainty (`None` or `np.ndarray`):
+            Uncalibrated uncertainty(MAD, median absolute deviation) coming from ensembling.
+    """
+    depth_np: np.ndarray
+    depth_colored: Image.Image
+    normal_np: np.ndarray
+    normal_colored: Image.Image
+    uncertainty: Union[None, np.ndarray]
+
+class DepthNormalEstimationPipeline(DiffusionPipeline):
+    # two hyper-parameters
+    latent_scale_factor = 0.18215
+
+    def __init__(self,
+                 unet:UNet2DConditionModel,
+                 vae:AutoencoderKL,
+                 scheduler:DDIMScheduler,
+                 image_encoder:CLIPVisionModelWithProjection,
+                 feature_extractor:CLIPImageProcessor,
+                 ):
+        super().__init__()
+            
+        self.register_modules(
+            unet=unet,
+            vae=vae,
+            scheduler=scheduler,
+            image_encoder=image_encoder,
+            feature_extractor=feature_extractor,
+        )
+        self.img_embed = None  
+
+    @torch.no_grad()
+    def __call__(self,
+                 input_image:Image,
+                 denosing_steps: int = 10,
+                 ensemble_size: int = 10,
+                 processing_res: int = 768,
+                 match_input_res:bool =True,
+                 batch_size:int = 0,
+                 domain: str = "indoor",
+                 color_map: str="Spectral",
+                 show_progress_bar:bool = True,
+                 ensemble_kwargs: Dict = None,
+                 ) -> DepthNormalPipelineOutput:
+        
+        # inherit from thea Diffusion Pipeline
+        device = self.device
+        input_size = input_image.size
+        
+        # adjust the input resolution.
+        if not match_input_res:
+            assert (
+                processing_res is not None                
+            )," Value Error: `resize_output_back` is only valid with "
+        
+        assert processing_res >=0
+        assert denosing_steps >=1
+        assert ensemble_size >=1
+
+        # --------------- Image Processing ------------------------
+        # Resize image
+        if processing_res >0:
+            input_image = resize_max_res(
+                input_image, max_edge_resolution=processing_res
+            )
+        
+        # Convert the image to RGB, to 1. reomve the alpha channel.
+        input_image = input_image.convert("RGB")
+        image = np.array(input_image)
+
+        # Normalize RGB Values.
+        rgb = np.transpose(image,(2,0,1))
+        rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
+        rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
+        rgb_norm = rgb_norm.to(device)
+
+        assert rgb_norm.min() >= -1.0 and rgb_norm.max() <= 1.0
+        
+        # ----------------- predicting depth -----------------
+        duplicated_rgb = torch.stack([rgb_norm] * ensemble_size)
+        single_rgb_dataset = TensorDataset(duplicated_rgb)
+        
+        # find the batch size
+        if batch_size>0:
+            _bs = batch_size
+        else:
+            _bs = 1
+
+        single_rgb_loader = DataLoader(single_rgb_dataset, batch_size=_bs, shuffle=False)
+        
+        # predicted the depth
+        depth_pred_ls = []
+        normal_pred_ls = []
+        
+        if show_progress_bar:
+            iterable_bar = tqdm(
+                single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
+            )
+        else:
+            iterable_bar = single_rgb_loader
+        
+        for batch in iterable_bar:
+            (batched_image, )= batch  # here the image is still around 0-1
+
+            depth_pred_raw, normal_pred_raw = self.single_infer(
+                input_rgb=batched_image,
+                num_inference_steps=denosing_steps,
+                domain=domain,
+                show_pbar=show_progress_bar,
+            )
+            depth_pred_ls.append(depth_pred_raw.detach().clone())
+            normal_pred_ls.append(normal_pred_raw.detach().clone())
+        
+        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze()
+        normal_preds = torch.concat(normal_pred_ls, axis=0).squeeze()
+        torch.cuda.empty_cache()  # clear vram cache for ensembling
+
+        # ----------------- Test-time ensembling -----------------
+        if ensemble_size > 1:
+            depth_pred, pred_uncert = ensemble_depths(
+                depth_preds, **(ensemble_kwargs or {})
+            )
+            normal_pred = ensemble_normals(normal_preds)
+        else:
+            depth_pred = depth_preds
+            normal_pred = normal_preds
+            pred_uncert = None
+
+        # ----------------- Post processing -----------------
+        # Scale prediction to [0, 1]
+        min_d = torch.min(depth_pred)
+        max_d = torch.max(depth_pred)
+        depth_pred = (depth_pred - min_d) / (max_d - min_d)
+        
+        # Convert to numpy
+        depth_pred = depth_pred.cpu().numpy().astype(np.float32)
+        normal_pred = normal_pred.cpu().numpy().astype(np.float32)
+
+        # Resize back to original resolution
+        if match_input_res:
+            pred_img = Image.fromarray(depth_pred)
+            pred_img = pred_img.resize(input_size)
+            depth_pred = np.asarray(pred_img)
+            normal_pred = cv2.resize(chw2hwc(normal_pred), input_size, interpolation = cv2.INTER_NEAREST)
+
+        # Clip output range: current size is the original size
+        depth_pred = depth_pred.clip(0, 1)
+        normal_pred = normal_pred.clip(-1, 1)
+    
+        # Colorize
+        depth_colored = colorize_depth_maps(
+            depth_pred, 0, 1, cmap=color_map
+        ).squeeze()  # [3, H, W], value in (0, 1)
+        depth_colored = (depth_colored * 255).astype(np.uint8)
+        depth_colored_hwc = chw2hwc(depth_colored)
+        depth_colored_img = Image.fromarray(depth_colored_hwc)
+
+        normal_colored = ((normal_pred + 1)/2 * 255).astype(np.uint8)
+        normal_colored_img = Image.fromarray(normal_colored)
+        
+        return DepthNormalPipelineOutput(
+            depth_np = depth_pred,
+            depth_colored = depth_colored_img,
+            normal_np = normal_pred,
+            normal_colored = normal_colored_img,
+            uncertainty=pred_uncert,
+        )
+    
+    def __encode_img_embed(self, rgb):
+        """
+        Encode clip embeddings for img
+        """
+        clip_image_mean = torch.as_tensor(self.feature_extractor.image_mean)[:,None,None].to(device=self.device, dtype=self.dtype)
+        clip_image_std = torch.as_tensor(self.feature_extractor.image_std)[:,None,None].to(device=self.device, dtype=self.dtype)
+
+        img_in_proc = TF.resize((rgb +1)/2, 
+            (self.feature_extractor.crop_size['height'], self.feature_extractor.crop_size['width']), 
+            interpolation=InterpolationMode.BICUBIC, 
+            antialias=True
+        )
+        # do the normalization in float32 to preserve precision
+        img_in_proc = ((img_in_proc.float() - clip_image_mean) / clip_image_std).to(self.dtype)        
+        img_embed = self.image_encoder(img_in_proc).image_embeds.unsqueeze(1).to(self.dtype)
+
+        self.img_embed = img_embed
+
+        
+    @torch.no_grad()
+    def single_infer(self,input_rgb:torch.Tensor,
+                     num_inference_steps:int,
+                     domain:str,
+                     show_pbar:bool,):
+
+        device = input_rgb.device
+
+        # Set timesteps: inherit from the diffuison pipeline
+        self.scheduler.set_timesteps(num_inference_steps, device=device) # here the numbers of the steps is only 10.
+        timesteps = self.scheduler.timesteps  # [T]
+        
+        # encode image
+        rgb_latent = self.encode_RGB(input_rgb)
+        
+        # Initial depth map (Guassian noise)
+        geo_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype).repeat(2,1,1,1)
+        rgb_latent = rgb_latent.repeat(2,1,1,1)
+
+        # Batched img embedding
+        if self.img_embed is None:
+            self.__encode_img_embed(input_rgb)
+        
+        batch_img_embed = self.img_embed.repeat(
+            (rgb_latent.shape[0], 1, 1)
+        )  # [B, 1, 768]
+
+        batch_img_embed = torch.cat((torch.zeros_like(batch_img_embed), batch_img_embed), dim=0)
+        rgb_latent = torch.cat((rgb_latent, rgb_latent), dim=0)
+
+        # hybrid switcher 
+        geo_class = torch.tensor([[0., 1.], [1, 0]], device=device, dtype=self.dtype)
+        geo_embedding = torch.cat([torch.sin(geo_class), torch.cos(geo_class)], dim=-1)
+        
+        if domain == "indoor":
+            domain_class = torch.tensor([[1., 0., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "outdoor":
+            domain_class = torch.tensor([[0., 1., 0]], device=device, dtype=self.dtype).repeat(2,1)
+        elif domain == "object":
+            domain_class = torch.tensor([[0., 0., 1]], device=device, dtype=self.dtype).repeat(2,1)
+        domain_embedding = torch.cat([torch.sin(domain_class), torch.cos(domain_class)], dim=-1)
+
+        class_embedding = torch.cat((geo_embedding, domain_embedding), dim=-1)
+
+        # Denoising loop
+        if show_pbar:
+            iterable = tqdm(
+                enumerate(timesteps),
+                total=len(timesteps),
+                leave=False,
+                desc=" " * 4 + "Diffusion denoising",
+            )
+        else:
+            iterable = enumerate(timesteps)
+
+        for i, t in iterable:
+            unet_input = torch.cat((rgb_latent, geo_latent.repeat(2,1,1,1)), dim=1)
+            # predict the noise residual
+            noise_pred = self.unet(unet_input, t.repeat(4), encoder_hidden_states=batch_img_embed, class_labels=class_embedding.repeat(2,1)).sample  
+            noise_pred_uncond, noise_pred_cond = noise_pred.chunk(2)
+            guidance_scale = 1.
+            noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
+        
+            # compute the previous noisy sample x_t -> x_t-1
+            geo_latent = self.scheduler.step(noise_pred, t, geo_latent).prev_sample
+
+        geo_latent = geo_latent
+        torch.cuda.empty_cache()
+
+        depth = self.decode_depth(geo_latent[0][None])
+        depth = torch.clip(depth, -1.0, 1.0)
+        depth = (depth + 1.0) / 2.0
+        
+        normal = self.decode_normal(geo_latent[1][None])
+        normal /= (torch.norm(normal, p=2, dim=1, keepdim=True)+1e-5)
+        normal *= -1.
+
+        return depth, normal
+        
+    
+    def encode_RGB(self, rgb_in: torch.Tensor) -> torch.Tensor:
+        """
+        Encode RGB image into latent.
+
+        Args:
+            rgb_in (`torch.Tensor`):
+                Input RGB image to be encoded.
+
+        Returns:
+            `torch.Tensor`: Image latent.
+        """
+
+        # encode
+        h = self.vae.encoder(rgb_in)
+
+        moments = self.vae.quant_conv(h)
+        mean, logvar = torch.chunk(moments, 2, dim=1)
+        # scale latent
+        rgb_latent = mean * self.latent_scale_factor
+        
+        return rgb_latent
+    
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode depth latent into depth map.
+
+        Args:
+            depth_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded depth map.
+        """
+
+        # scale latent
+        depth_latent = depth_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(depth_latent)
+        stacked = self.vae.decoder(z)
+        # mean of output channels
+        depth_mean = stacked.mean(dim=1, keepdim=True)
+        return depth_mean
+
+    def decode_normal(self, normal_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode normal latent into normal map.
+
+        Args:
+            normal_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded normal map.
+        """
+
+        # scale latent
+        normal_latent = normal_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(normal_latent)
+        normal = self.vae.decoder(z)
+        return normal
+
+

models/depth_pipeline.py

@@ -0,0 +1,310 @@
+# A reimplemented version in public environments by Xiao Fu and Mu Hu
+
+from typing import Any, Dict, Union
+
+import torch
+from torch.utils.data import DataLoader, TensorDataset
+import numpy as np
+from tqdm.auto import tqdm
+from PIL import Image
+from diffusers import (
+    DiffusionPipeline,
+    DDIMScheduler,
+    UNet2DConditionModel,
+    AutoencoderKL,
+)
+from diffusers.utils import BaseOutput
+from transformers import CLIPTextModel, CLIPTokenizer
+
+from utils.image_util import resize_max_res,chw2hwc,colorize_depth_maps
+from utils.colormap import kitti_colormap
+from utils.depth_ensemble import ensemble_depths
+from utils.batch_size import find_batch_size
+import cv2
+
+class DepthPipelineOutput(BaseOutput):
+    """
+    Output class for Marigold monocular depth prediction pipeline.
+
+    Args:
+        depth_np (`np.ndarray`):
+            Predicted depth map, with depth values in the range of [0, 1].
+        depth_colored (`PIL.Image.Image`):
+            Colorized depth map, with the shape of [3, H, W] and values in [0, 1].
+        uncertainty (`None` or `np.ndarray`):
+            Uncalibrated uncertainty(MAD, median absolute deviation) coming from ensembling.
+    """
+    depth_np: np.ndarray
+    depth_colored: Image.Image
+    uncertainty: Union[None, np.ndarray]
+
+class DepthEstimationPipeline(DiffusionPipeline):
+    # two hyper-parameters
+    latent_scale_factor = 0.18215
+
+    def __init__(self,
+                 unet:UNet2DConditionModel,
+                 vae:AutoencoderKL,
+                 scheduler:DDIMScheduler,
+                 text_encoder:CLIPTextModel,
+                 tokenizer:CLIPTokenizer,
+                 ):
+        super().__init__()
+            
+        self.register_modules(
+            unet=unet,
+            vae=vae,
+            scheduler=scheduler,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+        )
+        self.empty_text_embed = None  
+    
+    @torch.no_grad()
+    def __call__(self,
+                 input_image:Image,
+                 denosing_steps: int =10,
+                 ensemble_size: int =10,
+                 processing_res: int = 768,
+                 match_input_res:bool =True,
+                 batch_size:int = 0,
+                 color_map: str="Spectral",
+                 show_progress_bar:bool = True,
+                 ensemble_kwargs: Dict = None,
+                 ) -> DepthPipelineOutput:
+        
+        # inherit from thea Diffusion Pipeline
+        device = self.device
+        input_size = input_image.size
+        
+        # adjust the input resolution.
+        if not match_input_res:
+            assert (
+                processing_res is not None                
+            )," Value Error: `resize_output_back` is only valid with "
+        
+        assert processing_res >=0
+        assert denosing_steps >=1
+        assert ensemble_size >=1
+
+        # --------------- Image Processing ------------------------
+        # Resize image
+        if processing_res >0:
+            input_image = resize_max_res(
+                input_image, max_edge_resolution=processing_res
+            ) # resize image: for kitti is 231, 768
+        
+        # Convert the image to RGB, to 1. reomve the alpha channel.
+        input_image = input_image.convert("RGB")
+        image = np.array(input_image)
+
+        # Normalize RGB Values.
+        rgb = np.transpose(image,(2,0,1))
+        rgb_norm = rgb / 255.0 * 2.0 - 1.0 # [0, 255] -> [-1, 1]
+        rgb_norm = torch.from_numpy(rgb_norm).to(self.dtype)
+        rgb_norm = rgb_norm.to(device)
+
+        assert rgb_norm.min() >= -1.0 and rgb_norm.max() <= 1.0
+        
+        # ----------------- predicting depth -----------------
+        duplicated_rgb = torch.stack([rgb_norm] * ensemble_size)
+        single_rgb_dataset = TensorDataset(duplicated_rgb)
+        
+        # find the batch size
+        if batch_size>0:
+            _bs = batch_size
+        else:
+            _bs = 1
+
+        single_rgb_loader = DataLoader(single_rgb_dataset, batch_size=_bs, shuffle=False)
+        
+        # predicted the depth
+        depth_pred_ls = []
+        
+        if show_progress_bar:
+            iterable_bar = tqdm(
+                single_rgb_loader, desc=" " * 2 + "Inference batches", leave=False
+            )
+        else:
+            iterable_bar = single_rgb_loader
+        
+        for batch in iterable_bar:
+            (batched_image, )= batch  # here the image is still around 0-1
+
+            depth_pred_raw = self.single_infer(
+                input_rgb=batched_image,
+                num_inference_steps=denosing_steps,
+                show_pbar=show_progress_bar,
+            )
+            depth_pred_ls.append(depth_pred_raw.detach().clone())
+
+        depth_preds = torch.concat(depth_pred_ls, axis=0).squeeze() #(10,224,768)
+        torch.cuda.empty_cache()  # clear vram cache for ensembling
+
+        # ----------------- Test-time ensembling -----------------
+        if ensemble_size > 1:
+            depth_pred, pred_uncert = ensemble_depths(
+                depth_preds, **(ensemble_kwargs or {})
+            )
+        else:
+            depth_pred = depth_preds
+            pred_uncert = None
+
+        # ----------------- Post processing -----------------
+        # Scale prediction to [0, 1]
+        min_d = torch.min(depth_pred)
+        max_d = torch.max(depth_pred)
+        depth_pred = (depth_pred - min_d) / (max_d - min_d)
+        
+        # Convert to numpy
+        depth_pred = depth_pred.cpu().numpy().astype(np.float32)
+
+        # Resize back to original resolution
+        if match_input_res:
+            pred_img = Image.fromarray(depth_pred)
+            pred_img = pred_img.resize(input_size)
+            depth_pred = np.asarray(pred_img)
+
+        # Clip output range: current size is the original size
+        depth_pred = depth_pred.clip(0, 1)
+
+        # colorization using the KITTI Color Plan.
+        depth_pred_vis = depth_pred * 70
+        disp_vis = 400/(depth_pred_vis+1e-3)
+        disp_vis = disp_vis.clip(0,500)
+        
+        depth_color_pred =  kitti_colormap(disp_vis)
+    
+        # Colorize
+        depth_colored = colorize_depth_maps(
+            depth_pred, 0, 1, cmap=color_map
+        ).squeeze()  # [3, H, W], value in (0, 1)
+        depth_colored = (depth_colored * 255).astype(np.uint8)
+        depth_colored_hwc = chw2hwc(depth_colored)
+        depth_colored_img = Image.fromarray(depth_colored_hwc)
+
+        return DepthPipelineOutput(
+            depth_np = depth_pred,
+            depth_colored = depth_colored_img,
+            uncertainty=pred_uncert,
+        )
+    
+    def __encode_empty_text(self):
+        """
+        Encode text embedding for empty prompt
+        """
+        prompt = ""
+        text_inputs = self.tokenizer(
+            prompt,
+            padding="do_not_pad",
+            max_length=self.tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        )
+        text_input_ids = text_inputs.input_ids.to(self.text_encoder.device) #[1,2]
+        # print(text_input_ids.shape)
+        self.empty_text_embed = self.text_encoder(text_input_ids)[0].to(self.dtype) #[1,2,1024]
+
+        
+    @torch.no_grad()
+    def single_infer(self,input_rgb:torch.Tensor,
+                     num_inference_steps:int,
+                     show_pbar:bool,):
+
+        device = input_rgb.device
+
+        # Set timesteps: inherit from the diffuison pipeline
+        self.scheduler.set_timesteps(num_inference_steps, device=device) # here the numbers of the steps is only 10.
+        timesteps = self.scheduler.timesteps  # [T]
+        
+        # encode image
+        rgb_latent = self.encode_RGB(input_rgb) # 1/8 Resolution with a channel nums of 4. 
+        
+        # Initial depth map (Guassian noise)
+        depth_latent = torch.randn(rgb_latent.shape, device=device, dtype=self.dtype)  # [B, 4, H/8, W/8]
+        
+        # Batched empty text embedding
+        if self.empty_text_embed is None:
+            self.__encode_empty_text()
+            
+        batch_empty_text_embed = self.empty_text_embed.repeat(
+            (rgb_latent.shape[0], 1, 1)
+        )  # [B, 2, 1024]
+
+
+        # Denoising loop
+        if show_pbar:
+            iterable = tqdm(
+                enumerate(timesteps),
+                total=len(timesteps),
+                leave=False,
+                desc=" " * 4 + "Diffusion denoising",
+            )
+        else:
+            iterable = enumerate(timesteps)
+
+        for i, t in iterable:
+            unet_input = torch.cat([rgb_latent, depth_latent], dim=1)
+
+            # predict the noise residual
+            noise_pred = self.unet(
+                unet_input, t, encoder_hidden_states=batch_empty_text_embed
+            ).sample  # [B, 4, h, w]
+
+            # compute the previous noisy sample x_t -> x_t-1
+            depth_latent = self.scheduler.step(noise_pred, t, depth_latent).prev_sample
+
+        torch.cuda.empty_cache()
+
+        depth = self.decode_depth(depth_latent)
+        
+        depth = torch.clip(depth, -1.0, 1.0)
+        depth = (depth + 1.0) / 2.0
+
+        return depth
+        
+    
+    def encode_RGB(self, rgb_in: torch.Tensor) -> torch.Tensor:
+        """
+        Encode RGB image into latent.
+
+        Args:
+            rgb_in (`torch.Tensor`):
+                Input RGB image to be encoded.
+
+        Returns:
+            `torch.Tensor`: Image latent.
+        """
+
+        # encode
+        h = self.vae.encoder(rgb_in)
+
+        moments = self.vae.quant_conv(h)
+        mean, logvar = torch.chunk(moments, 2, dim=1)
+        # scale latent
+        rgb_latent = mean * self.latent_scale_factor
+        
+        return rgb_latent
+    
+    def decode_depth(self, depth_latent: torch.Tensor) -> torch.Tensor:
+        """
+        Decode depth latent into depth map.
+
+        Args:
+            depth_latent (`torch.Tensor`):
+                Depth latent to be decoded.
+
+        Returns:
+            `torch.Tensor`: Decoded depth map.
+        """
+
+        # scale latent
+        depth_latent = depth_latent / self.latent_scale_factor
+        # decode
+        z = self.vae.post_quant_conv(depth_latent)
+        stacked = self.vae.decoder(z)
+        # mean of output channels
+
+        depth_mean = stacked.mean(dim=1, keepdim=True)
+        return depth_mean
+

models/transformer_2d.py

@@ -0,0 +1,463 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Some modifications are reimplemented in public environments by Xiao Fu and Mu Hu
+
+from dataclasses import dataclass
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.models.embeddings import ImagePositionalEmbeddings
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, is_torch_version
+from models.attention import BasicTransformerBlock
+from diffusers.models.embeddings import PatchEmbed, PixArtAlphaTextProjection
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+    ):
+        super().__init__()
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d if USE_PEFT_BACKEND else LoRACompatibleConv
+        linear_cls = nn.Linear if USE_PEFT_BACKEND else LoRACompatibleLinear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        if self.is_input_continuous:
+            self.in_channels = in_channels
+
+            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+            if use_linear_projection:
+                self.proj_in = linear_cls(in_channels, inner_dim)
+            else:
+                self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+            self.height = sample_size
+            self.width = sample_size
+            self.num_vector_embeds = num_vector_embeds
+            self.num_latent_pixels = self.height * self.width
+
+            self.latent_image_embedding = ImagePositionalEmbeddings(
+                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+            )
+        elif self.is_input_patches:
+            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+            self.height = sample_size
+            self.width = sample_size
+
+            self.patch_size = patch_size
+            interpolation_scale = self.config.sample_size // 64  # => 64 (= 512 pixart) has interpolation scale 1
+            interpolation_scale = max(interpolation_scale, 1)
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+                interpolation_scale=interpolation_scale,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if self.is_input_continuous:
+            # TODO: should use out_channels for continuous projections
+            if use_linear_projection:
+                self.proj_out = linear_cls(inner_dim, in_channels)
+            else:
+                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            self.norm_out = nn.LayerNorm(inner_dim)
+            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+        elif self.is_input_patches and norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
+            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+        elif self.is_input_patches and norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+            self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *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).
+            attention_mask ( `torch.Tensor`, *optional*):
+                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.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformer_2d.Transformer2DModelOutput`] is returned, otherwise a
+            `tuple` where the first element is the sample tensor.
+        """
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # 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 and attention_mask.ndim == 2:
+            # 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(hidden_states.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 and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # Retrieve lora scale.
+        lora_scale = cross_attention_kwargs.get("scale", 1.0) if cross_attention_kwargs is not None else 1.0
+
+        # 1. Input
+        if self.is_input_continuous:
+            batch, _, height, width = hidden_states.shape
+            residual = hidden_states
+
+            hidden_states = self.norm(hidden_states)
+            if not self.use_linear_projection:
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            else:
+                inner_dim = hidden_states.shape[1]
+                hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+                hidden_states = (
+                    self.proj_in(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_in(hidden_states)
+                )
+
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            hidden_states = self.pos_embed(hidden_states)
+
+            if self.adaln_single is not None:
+                if self.use_additional_conditions and added_cond_kwargs is None:
+                    raise ValueError(
+                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                    )
+                batch_size = hidden_states.shape[0]
+                timestep, embedded_timestep = self.adaln_single(
+                    timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+                )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            if not self.use_linear_projection:
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+            else:
+                hidden_states = (
+                    self.proj_out(hidden_states, scale=lora_scale)
+                    if not USE_PEFT_BACKEND
+                    else self.proj_out(hidden_states)
+                )
+                hidden_states = hidden_states.reshape(batch, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+
+            output = hidden_states + residual
+        elif self.is_input_vectorized:
+            hidden_states = self.norm_out(hidden_states)
+            logits = self.out(hidden_states)
+            # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+            logits = logits.permute(0, 2, 1)
+
+            # log(p(x_0))
+            output = F.log_softmax(logits.double(), dim=1).float()
+
+        if self.is_input_patches:
+            if self.config.norm_type != "ada_norm_single":
+                conditioning = self.transformer_blocks[0].norm1.emb(
+                    timestep, class_labels, hidden_dtype=hidden_states.dtype
+                )
+                shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+                hidden_states = self.proj_out_2(hidden_states)
+            elif self.config.norm_type == "ada_norm_single":
+                shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states)
+                # Modulation
+                hidden_states = hidden_states * (1 + scale) + shift
+                hidden_states = self.proj_out(hidden_states)
+                hidden_states = hidden_states.squeeze(1)
+
+            # unpatchify
+            if self.adaln_single is None:
+                height = width = int(hidden_states.shape[1] ** 0.5)
+            hidden_states = hidden_states.reshape(
+                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+            )
+            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+            output = hidden_states.reshape(
+                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+            )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

models/unet_2d_condition.py

@@ -0,0 +1,1214 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Some modifications are reimplemented in public environments by Xiao Fu and Mu Hu
+
+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.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    PositionNet,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+
+from models.unet_2d_blocks import (
+    UNetMidBlock2D,
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    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
+
+
+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":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                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,
+                cross_attention_dim=cross_attention_dim[-1],
+                attention_head_dim=attention_head_dim[-1],
+                resnet_groups=norm_num_groups,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                skip_time_act=resnet_skip_time_act,
+                only_cross_attention=mid_block_only_cross_attention,
+                cross_attention_norm=cross_attention_norm,
+            )
+        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
+
+        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 fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    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,)
+        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)