add mvdiffusion models

mvdiffusion/models/transformer_mv2d_image.py

@@ -0,0 +1,1029 @@
+# 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.
+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 BaseOutput, deprecate
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.attention import FeedForward, AdaLayerNorm, AdaLayerNormZero, Attention
+from diffusers.models.embeddings import PatchEmbed
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils.import_utils import is_xformers_available
+
+from einops import rearrange, repeat
+import pdb
+import random
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+def my_repeat(tensor, num_repeats):
+    """
+    Repeat a tensor along a given dimension
+    """
+    if len(tensor.shape) == 3:
+        return repeat(tensor,  "b d c -> (b v) d c", v=num_repeats)
+    elif len(tensor.shape) == 4:
+        return repeat(tensor,  "a b d c -> (a v) b d c", v=num_repeats)
+    
+
+@dataclass
+class TransformerMV2DModelOutput(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 TransformerMV2DModel(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.
+    """
+
+    @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,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        num_views: int = 1,
+        cd_attention_last: bool=False,
+        cd_attention_mid: bool=False,
+        multiview_attention: bool=True,
+        sparse_mv_attention: bool = False,
+        mvcd_attention: bool=False
+    ):
+        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
+
+        # 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 = LoRACompatibleLinear(in_channels, inner_dim)
+            else:
+                self.proj_in = LoRACompatibleConv(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
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicMVTransformerBlock(
+                    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,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    num_views=num_views,
+                    cd_attention_last=cd_attention_last,
+                    cd_attention_mid=cd_attention_mid,
+                    multiview_attention=multiview_attention,
+                    sparse_mv_attention=sparse_mv_attention,
+                    mvcd_attention=mvcd_attention
+                )
+                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 = LoRACompatibleLinear(inner_dim, in_channels)
+            else:
+                self.proj_out = LoRACompatibleConv(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:
+            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)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = 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`.
+            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)
+
+        # 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)
+                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)
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            hidden_states = self.pos_embed(hidden_states)
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            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)
+            else:
+                hidden_states = 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()
+        elif self.is_input_patches:
+            # TODO: cleanup!
+            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)
+
+            # unpatchify
+            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 TransformerMV2DModelOutput(sample=output)
+
+
+@maybe_allow_in_graph
+class BasicMVTransformerBlock(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.
+        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.
+        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.
+    """
+
+    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",
+        final_dropout: bool = False,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        sparse_mv_attention: bool = False,
+        mvcd_attention: bool = False
+    ):
+        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"
+
+        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}."
+            )
+
+        # 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)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+
+        self.multiview_attention = multiview_attention
+        self.sparse_mv_attention = sparse_mv_attention
+        self.mvcd_attention = mvcd_attention
+        
+        self.attn1 = CustomAttention(
+            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,
+            processor=MVAttnProcessor()
+        )
+
+        # 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.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=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,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+        self.num_views = num_views
+
+        self.cd_attention_last = cd_attention_last
+
+        if self.cd_attention_last:
+            # Joint task -Attn
+            self.attn_joint_last = 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,
+                processor=JointAttnProcessor()
+            )
+            nn.init.zeros_(self.attn_joint_last.to_out[0].weight.data)
+            self.norm_joint_last = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+
+        self.cd_attention_mid = cd_attention_mid
+
+        if self.cd_attention_mid:
+            print("cross-domain attn in the middle")
+            # Joint task -Attn
+            self.attn_joint_mid = 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,
+                processor=JointAttnProcessor()
+            )
+            nn.init.zeros_(self.attn_joint_mid.to_out[0].weight.data)
+            self.norm_joint_mid = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # 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,
+    ):
+        assert attention_mask is None # not supported yet
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 1. Self-Attention
+        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
+            )
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            num_views=self.num_views,
+            multiview_attention=self.multiview_attention,
+            sparse_mv_attention=self.sparse_mv_attention,
+            mvcd_attention=self.mvcd_attention,
+            **cross_attention_kwargs,
+        )
+
+
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+
+        # joint attention twice
+        if self.cd_attention_mid:
+            norm_hidden_states = (
+                self.norm_joint_mid(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_joint_mid(hidden_states)
+            )
+            hidden_states = self.attn_joint_mid(norm_hidden_states) + hidden_states
+
+        # 2. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(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
+
+        # 3. Feed-forward
+        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._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [self.ff(hid_slice) for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = ff_output + hidden_states
+
+        if self.cd_attention_last:
+            norm_hidden_states = (
+                self.norm_joint_last(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_joint_last(hidden_states)
+            )
+            hidden_states = self.attn_joint_last(norm_hidden_states) + hidden_states
+
+        return hidden_states
+    
+
+class CustomAttention(Attention):
+    def set_use_memory_efficient_attention_xformers(
+        self, use_memory_efficient_attention_xformers: bool, *args, **kwargs
+    ):
+        processor = XFormersMVAttnProcessor()
+        self.set_processor(processor)
+        # print("using xformers attention processor")
+
+
+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 MVAttnProcessor:
+    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_views=1,
+        multiview_attention=True
+    ):
+        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)
+
+        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)
+
+        # print('query', query.shape, 'key', key.shape, 'value', value.shape)
+        #([bx4, 1024, 320]) key torch.Size([bx4, 1024, 320]) value torch.Size([bx4, 1024, 320])
+        # pdb.set_trace()
+        # multi-view self-attention
+        if multiview_attention:
+            if num_views <= 6:
+                # after use xformer; possible to train with 6 views
+                key = rearrange(key, "(b t) d c -> b (t d) c", t=num_views).repeat_interleave(num_views, dim=0)
+                value = rearrange(value, "(b t) d c -> b (t d) c", t=num_views).repeat_interleave(num_views, dim=0)
+            else:# apply sparse attention
+                pass
+                # print("use sparse attention")  
+                # # seems that the sparse random sampling cause problems
+                # # don't use random sampling, just fix the indexes
+                # onekey = rearrange(key, "(b t) d c -> b t d c", t=num_views) 
+                # onevalue = rearrange(value, "(b t) d c -> b t d c", t=num_views)
+                # allkeys = []
+                # allvalues = []
+                # all_indexes = {
+                #     0 : [0, 2, 3, 4],
+                #     1: [0, 1, 3, 5],
+                #     2: [0, 2, 3, 4],
+                #     3: [0, 2, 3, 4],
+                #     4: [0, 2, 3, 4],
+                #     5: [0, 1, 3, 5]
+                # }
+                # for jj in range(num_views):
+                #     # valid_index = [x for x in range(0, num_views) if x!= jj]
+                #     # indexes = random.sample(valid_index, 3) + [jj] + [0]
+                #     indexes = all_indexes[jj]
+
+                #     indexes = torch.tensor(indexes).long().to(key.device)
+                #     allkeys.append(onekey[:, indexes])
+                #     allvalues.append(onevalue[:, indexes])
+                # keys = torch.stack(allkeys, dim=1) # checked, should be dim=1
+                # values = torch.stack(allvalues, dim=1) 
+                # key = rearrange(keys, 'b t f d c -> (b t) (f d) c')
+                # value = rearrange(values, 'b t f d c -> (b t) (f d) 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()
+        
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        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
+
+
+class XFormersMVAttnProcessor:
+    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_views=1.,
+        multiview_attention=True,
+        sparse_mv_attention=False,
+        mvcd_attention=False,
+    ):
+        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_raw = attn.to_k(encoder_hidden_states)
+        value_raw = attn.to_v(encoder_hidden_states)
+
+        # print('query', query.shape, 'key', key.shape, 'value', value.shape)
+        #([bx4, 1024, 320]) key torch.Size([bx4, 1024, 320]) value torch.Size([bx4, 1024, 320])
+        # pdb.set_trace()
+        # multi-view self-attention
+        if multiview_attention:
+            if not sparse_mv_attention:
+                key = my_repeat(rearrange(key_raw, "(b t) d c -> b (t d) c", t=num_views), num_views)
+                value = my_repeat(rearrange(value_raw, "(b t) d c -> b (t d) c", t=num_views), num_views)
+            else:
+                key_front = my_repeat(rearrange(key_raw, "(b t) d c -> b t d c", t=num_views)[:, 0, :, :], num_views) # [(b t), d, c]
+                value_front = my_repeat(rearrange(value_raw, "(b t) d c -> b t d c", t=num_views)[:, 0, :, :], num_views)
+                key = torch.cat([key_front, key_raw], dim=1) # shape (b t) (2 d) c
+                value = torch.cat([value_front, value_raw], dim=1)
+
+            if mvcd_attention:
+                # memory efficient, cross domain attention
+                key_0, key_1 = torch.chunk(key_raw, dim=0, chunks=2)  # keys shape (b t) d c
+                value_0, value_1 = torch.chunk(value_raw, dim=0, chunks=2)
+                key_cross = torch.concat([key_1, key_0], dim=0)
+                value_cross = torch.concat([value_1, value_0], dim=0) #  shape (b t) d c
+                key = torch.cat([key, key_cross], dim=1)
+                value = torch.cat([value, value_cross], dim=1)  #  shape (b t) (t+1 d) c
+        else:
+            # print("don't use multiview attention.")
+            key = key_raw
+            value = value_raw
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        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
+
+
+
+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_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
+
+
+class JointAttnProcessor:
+    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)
+
+
+        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_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()
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        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
+    
+

mvdiffusion/models/transformer_mv2d_rowwise.py

@@ -0,0 +1,978 @@
+# 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.
+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 BaseOutput, deprecate
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.attention import FeedForward, AdaLayerNorm, AdaLayerNormZero, Attention
+from diffusers.models.embeddings import PatchEmbed
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils.import_utils import is_xformers_available
+
+from einops import rearrange
+import pdb
+import random
+import math
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+@dataclass
+class TransformerMV2DModelOutput(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 TransformerMV2DModel(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.
+    """
+
+    @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,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        num_views: int = 1,
+        cd_attention_last: bool=False,
+        cd_attention_mid: bool=False,
+        multiview_attention: bool=True,
+        sparse_mv_attention: bool = True, # not used
+        mvcd_attention: bool=False
+    ):
+        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
+
+        # 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 = LoRACompatibleLinear(in_channels, inner_dim)
+            else:
+                self.proj_in = LoRACompatibleConv(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
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicMVTransformerBlock(
+                    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,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    num_views=num_views,
+                    cd_attention_last=cd_attention_last,
+                    cd_attention_mid=cd_attention_mid,
+                    multiview_attention=multiview_attention,
+                    mvcd_attention=mvcd_attention
+                )
+                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 = LoRACompatibleLinear(inner_dim, in_channels)
+            else:
+                self.proj_out = LoRACompatibleConv(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:
+            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)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = 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`.
+            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)
+
+        # 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)
+                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)
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            hidden_states = self.pos_embed(hidden_states)
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            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)
+            else:
+                hidden_states = 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()
+        elif self.is_input_patches:
+            # TODO: cleanup!
+            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)
+
+            # unpatchify
+            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 TransformerMV2DModelOutput(sample=output)
+
+
+@maybe_allow_in_graph
+class BasicMVTransformerBlock(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.
+        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.
+        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.
+    """
+
+    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",
+        final_dropout: bool = False,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        mvcd_attention: bool = False,
+        rowwise_attention: 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"
+
+        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}."
+            )
+
+        # 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)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+
+        self.multiview_attention = multiview_attention
+        self.mvcd_attention = mvcd_attention
+        self.rowwise_attention = multiview_attention and rowwise_attention
+
+        # rowwise multiview attention
+     
+        print('INFO: using row wise attention...')
+    
+        self.attn1 = CustomAttention(
+            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,
+            processor=MVAttnProcessor()
+        )
+
+        # 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.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=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,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+        self.num_views = num_views
+
+        self.cd_attention_last = cd_attention_last
+
+        if self.cd_attention_last:
+            # Joint task -Attn
+            self.attn_joint = 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,
+                processor=JointAttnProcessor()
+            )
+            nn.init.zeros_(self.attn_joint.to_out[0].weight.data)
+            self.norm_joint = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+
+        self.cd_attention_mid = cd_attention_mid
+
+        if self.cd_attention_mid:
+            print("joint twice")
+            # Joint task -Attn
+            self.attn_joint_twice = 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,
+                processor=JointAttnProcessor()
+            )
+            nn.init.zeros_(self.attn_joint_twice.to_out[0].weight.data)
+            self.norm_joint_twice = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # 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,
+    ):
+        assert attention_mask is None # not supported yet
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 1. Self-Attention
+        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
+            )
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            multiview_attention=self.multiview_attention,
+            mvcd_attention=self.mvcd_attention,
+            num_views=self.num_views,
+            **cross_attention_kwargs,
+        )
+
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+
+        # joint attention twice
+        if self.cd_attention_mid:
+            norm_hidden_states = (
+                self.norm_joint_twice(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_joint_twice(hidden_states)
+            )
+            hidden_states = self.attn_joint_twice(norm_hidden_states) + hidden_states
+
+        # 2. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(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
+
+        # 3. Feed-forward
+        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._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [self.ff(hid_slice) for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = ff_output + hidden_states
+
+        if self.cd_attention_last:
+            norm_hidden_states = (
+                self.norm_joint(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_joint(hidden_states)
+            )
+            hidden_states = self.attn_joint(norm_hidden_states) + hidden_states
+
+        return hidden_states
+    
+
+class CustomAttention(Attention):
+    def set_use_memory_efficient_attention_xformers(
+        self, use_memory_efficient_attention_xformers: bool, *args, **kwargs
+    ):
+        processor = XFormersMVAttnProcessor()
+        self.set_processor(processor)
+        # print("using xformers attention processor")
+
+
+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 MVAttnProcessor:
+    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_views=1,
+        multiview_attention=True
+    ):
+        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
+        )
+        height = int(math.sqrt(sequence_length)) 
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        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)
+
+        # print('query', query.shape, 'key', key.shape, 'value', value.shape)
+        #([bx4, 1024, 320]) key torch.Size([bx4, 1024, 320]) value torch.Size([bx4, 1024, 320])
+        # pdb.set_trace()
+        # multi-view self-attention
+        key = rearrange(key, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) 
+        value = rearrange(value, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height)
+        query = rearrange(query, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) # torch.Size([192, 384, 320])
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+        
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        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)
+        hidden_states = rearrange(hidden_states, "(b h) (v w) c -> (b v) (h w) c", v=num_views, h=height)
+        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
+
+
+class XFormersMVAttnProcessor:
+    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_views=1,
+        multiview_attention=True,
+        mvcd_attention=False,
+    ):
+        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
+        )
+        height = int(math.sqrt(sequence_length)) 
+        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:
+            print('Warning: using group norm, pay attention to use it in row-wise attention')
+            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_raw = attn.to_k(encoder_hidden_states)
+        value_raw = attn.to_v(encoder_hidden_states)
+
+        # print('query', query.shape, 'key', key.shape, 'value', value.shape)
+        # pdb.set_trace()
+        
+        key = rearrange(key_raw, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) 
+        value = rearrange(value_raw, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height)
+        query = rearrange(query, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) # torch.Size([192, 384, 320])
+        if mvcd_attention:
+            # memory efficient, cross domain attention
+            key_0, key_1 = torch.chunk(key_raw, dim=0, chunks=2)  # keys shape (b t) d c
+            value_0, value_1 = torch.chunk(value_raw, dim=0, chunks=2)
+            key_cross = torch.concat([key_1, key_0], dim=0)
+            value_cross = torch.concat([value_1, value_0], dim=0) #  shape (b t) d c
+            key = torch.cat([key, key_cross], dim=1)
+            value = torch.cat([value, value_cross], dim=1)  #  shape (b t) (t+1 d) c
+
+
+        query = attn.head_to_batch_dim(query) # torch.Size([960, 384, 64])
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        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)
+        # print(hidden_states.shape)
+        hidden_states = rearrange(hidden_states, "(b h) (v w) c -> (b v) (h w) c", v=num_views, h=height)
+        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
+
+
+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_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
+
+
+class JointAttnProcessor:
+    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)
+
+
+        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_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()
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        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

mvdiffusion/models/transformer_mv2d_self_rowwise.py

@@ -0,0 +1,1038 @@
+# 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.
+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 BaseOutput, deprecate
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.attention import FeedForward, AdaLayerNorm, AdaLayerNormZero, Attention
+from diffusers.models.embeddings import PatchEmbed
+from diffusers.models.lora import LoRACompatibleConv, LoRACompatibleLinear
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.utils.import_utils import is_xformers_available
+
+from einops import rearrange
+import pdb
+import random
+import math
+
+
+if is_xformers_available():
+    import xformers
+    import xformers.ops
+else:
+    xformers = None
+
+
+@dataclass
+class TransformerMV2DModelOutput(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 TransformerMV2DModel(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.
+    """
+
+    @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,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",
+        norm_elementwise_affine: bool = True,
+        num_views: int = 1,
+        cd_attention_mid: bool=False,
+        cd_attention_last: bool=False,
+        multiview_attention: bool=True,
+        sparse_mv_attention: bool = True, # not used
+        mvcd_attention: bool=False,
+        use_dino: bool=False
+    ):
+        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
+
+        # 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 = LoRACompatibleLinear(in_channels, inner_dim)
+            else:
+                self.proj_in = LoRACompatibleConv(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
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicMVTransformerBlock(
+                    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,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    num_views=num_views,
+                    cd_attention_last=cd_attention_last,
+                    cd_attention_mid=cd_attention_mid,
+                    multiview_attention=multiview_attention,
+                    mvcd_attention=mvcd_attention,
+                    use_dino=use_dino
+                )
+                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 = LoRACompatibleLinear(inner_dim, in_channels)
+            else:
+                self.proj_out = LoRACompatibleConv(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:
+            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)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        dino_feature: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = 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`.
+            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)
+
+        # 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)
+                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)
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            hidden_states = self.pos_embed(hidden_states)
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            hidden_states = block(
+                hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                dino_feature=dino_feature,
+                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)
+            else:
+                hidden_states = 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()
+        elif self.is_input_patches:
+            # TODO: cleanup!
+            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)
+
+            # unpatchify
+            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 TransformerMV2DModelOutput(sample=output)
+
+
+@maybe_allow_in_graph
+class BasicMVTransformerBlock(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.
+        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.
+        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.
+    """
+
+    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",
+        final_dropout: bool = False,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        mvcd_attention: bool = False,
+        rowwise_attention: bool = True,
+        use_dino: bool = False
+    ):
+        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"
+
+        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}."
+            )
+
+        # 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)
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+
+        self.multiview_attention = multiview_attention
+        self.mvcd_attention = mvcd_attention
+        self.cd_attention_mid = cd_attention_mid
+        self.rowwise_attention = multiview_attention and rowwise_attention
+
+        if mvcd_attention and (not cd_attention_mid):
+        # add cross domain attn to self attn
+            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,
+                processor=JointAttnProcessor()
+            )
+        else:
+            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 if only_cross_attention else None,
+                upcast_attention=upcast_attention
+            )
+        # 1.1 rowwise multiview attention
+        if self.rowwise_attention:
+            # print('INFO: using self+row_wise mv attention...')
+            self.norm_mv = (
+                    AdaLayerNorm(dim, num_embeds_ada_norm)
+                    if self.use_ada_layer_norm
+                    else nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+            )
+            self.attn_mv = CustomAttention(
+                    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,
+                    processor=MVAttnProcessor() 
+                )
+            nn.init.zeros_(self.attn_mv.to_out[0].weight.data)
+        else:
+            self.norm_mv = None
+            self.attn_mv = None
+
+        # # 1.2 rowwise cross-domain attn
+        # if mvcd_attention:
+        #     self.attn_joint = 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,
+        #         processor=JointAttnProcessor()
+        #     )
+        #     nn.init.zeros_(self.attn_joint.to_out[0].weight.data)
+        #     self.norm_joint = AdaLayerNorm(dim, num_embeds_ada_norm) if self.use_ada_layer_norm else nn.LayerNorm(dim)
+        # else:
+        #     self.attn_joint = None
+        #     self.norm_joint = None
+            
+        # 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.
+            self.norm2 = (
+                AdaLayerNorm(dim, num_embeds_ada_norm)
+                if self.use_ada_layer_norm
+                else nn.LayerNorm(dim, elementwise_affine=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,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine)
+        self.ff = FeedForward(dim, dropout=dropout, activation_fn=activation_fn, final_dropout=final_dropout)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+        self.num_views = num_views
+
+       
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int):
+        # 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,
+        dino_feature: Optional[torch.FloatTensor] = None
+    ):
+        assert attention_mask is None # not supported yet
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 1. Self-Attention
+        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
+            )
+        else:
+            norm_hidden_states = self.norm1(hidden_states)
+
+        cross_attention_kwargs = cross_attention_kwargs if cross_attention_kwargs is not None else {}
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            # multiview_attention=self.multiview_attention,
+            # mvcd_attention=self.mvcd_attention,
+            **cross_attention_kwargs,
+        )
+
+
+        if self.use_ada_layer_norm_zero:
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = attn_output + hidden_states
+        
+        # import pdb;pdb.set_trace()
+        # 1.1 row wise multiview attention
+        if self.rowwise_attention:
+            norm_hidden_states = (
+                self.norm_mv(hidden_states, timestep) if self.use_ada_layer_norm else self.norm_mv(hidden_states)
+            )
+            attn_output = self.attn_mv(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                attention_mask=attention_mask,
+                num_views=self.num_views,
+                multiview_attention=self.multiview_attention,
+                cd_attention_mid=self.cd_attention_mid,
+                **cross_attention_kwargs,
+                )
+            hidden_states = attn_output + hidden_states 
+            
+    
+        # 2. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = (
+                self.norm2(hidden_states, timestep) if self.use_ada_layer_norm else self.norm2(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
+
+        # 3. Feed-forward
+        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._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            if norm_hidden_states.shape[self._chunk_dim] % self._chunk_size != 0:
+                raise ValueError(
+                    f"`hidden_states` dimension to be chunked: {norm_hidden_states.shape[self._chunk_dim]} has to be divisible by chunk size: {self._chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+                )
+
+            num_chunks = norm_hidden_states.shape[self._chunk_dim] // self._chunk_size
+            ff_output = torch.cat(
+                [self.ff(hid_slice) for hid_slice in norm_hidden_states.chunk(num_chunks, dim=self._chunk_dim)],
+                dim=self._chunk_dim,
+            )
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.use_ada_layer_norm_zero:
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = ff_output + hidden_states
+
+        return hidden_states
+    
+
+class CustomAttention(Attention):
+    def set_use_memory_efficient_attention_xformers(
+        self, use_memory_efficient_attention_xformers: bool, *args, **kwargs
+    ):
+        processor = XFormersMVAttnProcessor()
+        self.set_processor(processor)
+        # print("using xformers attention processor")
+
+
+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 MVAttnProcessor:
+    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_views=1,
+        cd_attention_mid=False
+    ):
+        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
+        )
+        height = int(math.sqrt(sequence_length)) 
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        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)
+
+        # print('query', query.shape, 'key', key.shape, 'value', value.shape)
+        #([bx4, 1024, 320]) key torch.Size([bx4, 1024, 320]) value torch.Size([bx4, 1024, 320])
+        # pdb.set_trace()
+        # multi-view self-attention
+        def transpose(tensor):
+            tensor = rearrange(tensor, "(b v) (h w) c -> b v h w c", v=num_views, h=height)
+            tensor_0, tensor_1 = torch.chunk(tensor, dim=0, chunks=2)  # b v h w c
+            tensor = torch.cat([tensor_0, tensor_1], dim=3)  # b v h 2w c
+            tensor = rearrange(tensor, "b v h w c -> (b h) (v w) c", v=num_views, h=height)
+            return tensor
+
+        if cd_attention_mid:
+            key = transpose(key)
+            value = transpose(value)
+            query = transpose(query)
+        else:
+            key = rearrange(key, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) 
+            value = rearrange(value, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height)
+            query = rearrange(query, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) # torch.Size([192, 384, 320])
+
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+        
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        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 cd_attention_mid:
+            hidden_states = rearrange(hidden_states, "(b h) (v w) c -> b v h w c", v=num_views, h=height)
+            hidden_states_0, hidden_states_1 = torch.chunk(hidden_states, dim=3, chunks=2)  # b v h w c
+            hidden_states = torch.cat([hidden_states_0, hidden_states_1], dim=0)  # 2b v h w c
+            hidden_states = rearrange(hidden_states, "b v h w c -> (b v) (h w) c", v=num_views, h=height) 
+        else:
+            hidden_states = rearrange(hidden_states, "(b h) (v w) c -> (b v) (h w) c", v=num_views, h=height)
+        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
+
+
+class XFormersMVAttnProcessor:
+    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_views=1,
+        multiview_attention=True,
+        cd_attention_mid=False
+    ):
+        # print(num_views)
+        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
+        )
+        height = int(math.sqrt(sequence_length)) 
+        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:
+            print('Warning: using group norm, pay attention to use it in row-wise attention')
+            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_raw = attn.to_k(encoder_hidden_states)
+        value_raw = attn.to_v(encoder_hidden_states)
+
+        # print('query', query.shape, 'key', key.shape, 'value', value.shape)
+        # pdb.set_trace()
+        def transpose(tensor):
+            tensor = rearrange(tensor, "(b v) (h w) c -> b v h w c", v=num_views, h=height)
+            tensor_0, tensor_1 = torch.chunk(tensor, dim=0, chunks=2)  # b v h w c
+            tensor = torch.cat([tensor_0, tensor_1], dim=3)  # b v h 2w c
+            tensor = rearrange(tensor, "b v h w c -> (b h) (v w) c", v=num_views, h=height)
+            return tensor
+        # print(mvcd_attention)
+        # import pdb;pdb.set_trace()
+        if cd_attention_mid:
+            key = transpose(key_raw)
+            value = transpose(value_raw)
+            query = transpose(query)
+        else:
+            key = rearrange(key_raw, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) 
+            value = rearrange(value_raw, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height)
+            query = rearrange(query, "(b v) (h w) c -> (b h) (v w) c", v=num_views, h=height) # torch.Size([192, 384, 320])
+
+
+        query = attn.head_to_batch_dim(query) # torch.Size([960, 384, 64])
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        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 cd_attention_mid:
+            hidden_states = rearrange(hidden_states, "(b h) (v w) c -> b v h w c", v=num_views, h=height)
+            hidden_states_0, hidden_states_1 = torch.chunk(hidden_states, dim=3, chunks=2)  # b v h w c
+            hidden_states = torch.cat([hidden_states_0, hidden_states_1], dim=0)  # 2b v h w c
+            hidden_states = rearrange(hidden_states, "b v h w c -> (b v) (h w) c", v=num_views, h=height) 
+        else:
+            hidden_states = rearrange(hidden_states, "(b h) (v w) c -> (b v) (h w) c", v=num_views, h=height)
+        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
+
+
+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
+
+        def transpose(tensor):
+            tensor_0, tensor_1 = torch.chunk(tensor, dim=0, chunks=2)  # bv hw c
+            tensor = torch.cat([tensor_0, tensor_1], dim=1)  # bv 2hw c
+            return tensor 
+        key  = transpose(key)
+        value = transpose(value)
+        query = transpose(query)
+        # from icecream import ic
+        # ic(key.shape, value.shape, query.shape)
+        # import pdb;pdb.set_trace()
+        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)
+        hidden_states_normal, hidden_states_color = torch.chunk(hidden_states, dim=1, chunks=2)
+        hidden_states = torch.cat([hidden_states_normal, hidden_states_color], dim=0)  # 2bv hw c
+        
+        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
+
+
+class JointAttnProcessor:
+    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)
+
+
+        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
+
+        def transpose(tensor):
+            tensor_0, tensor_1 = torch.chunk(tensor, dim=0, chunks=2)  # bv hw c
+            tensor = torch.cat([tensor_0, tensor_1], dim=1)  # bv 2hw c
+            return tensor 
+        key  = transpose(key)
+        value = transpose(value)
+        query = transpose(query)
+
+        
+        query = attn.head_to_batch_dim(query).contiguous()
+        key = attn.head_to_batch_dim(key).contiguous()
+        value = attn.head_to_batch_dim(value).contiguous()
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        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)
+
+        hidden_states = torch.cat([hidden_states[:, 0], hidden_states[:, 1]], dim=0)  # 2bv hw c
+        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

mvdiffusion/models/unet_mv2d_blocks.py

@@ -0,0 +1,971 @@
+# 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.
+from typing import Any, Dict, Optional, Tuple
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.utils import is_torch_version, logging
+from diffusers.models.normalization import AdaGroupNorm
+from diffusers.models.attention_processor import Attention, AttnAddedKVProcessor, AttnAddedKVProcessor2_0
+from diffusers.models.dual_transformer_2d import DualTransformer2DModel
+from diffusers.models.resnet import Downsample2D, FirDownsample2D, FirUpsample2D, KDownsample2D, KUpsample2D, ResnetBlock2D, Upsample2D
+
+from diffusers.models.unets.unet_2d_blocks import DownBlock2D, ResnetDownsampleBlock2D, AttnDownBlock2D, CrossAttnDownBlock2D, SimpleCrossAttnDownBlock2D, SkipDownBlock2D, AttnSkipDownBlock2D, DownEncoderBlock2D, AttnDownEncoderBlock2D, KDownBlock2D, KCrossAttnDownBlock2D
+from diffusers.models.unets.unet_2d_blocks import UpBlock2D, ResnetUpsampleBlock2D, CrossAttnUpBlock2D, SimpleCrossAttnUpBlock2D, AttnUpBlock2D, SkipUpBlock2D, AttnSkipUpBlock2D, UpDecoderBlock2D, AttnUpDecoderBlock2D, KUpBlock2D, KCrossAttnUpBlock2D
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+def get_down_block(
+    down_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    temb_channels,
+    add_downsample,
+    resnet_eps,
+    resnet_act_fn,
+    transformer_layers_per_block=1,
+    num_attention_heads=None,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    downsample_padding=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    resnet_skip_time_act=False,
+    resnet_out_scale_factor=1.0,
+    cross_attention_norm=None,
+    attention_head_dim=None,
+    downsample_type=None,
+    num_views=1,
+    cd_attention_last: bool = False,
+    cd_attention_mid: bool = False,
+    multiview_attention: bool = True,
+    sparse_mv_attention: bool = False,
+    selfattn_block: str = "custom",
+    mvcd_attention: bool=False,
+    use_dino: bool = False
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_down_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    down_block_type = down_block_type[7:] if down_block_type.startswith("UNetRes") else down_block_type
+    if down_block_type == "DownBlock2D":
+        return DownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "ResnetDownsampleBlock2D":
+        return ResnetDownsampleBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            skip_time_act=resnet_skip_time_act,
+            output_scale_factor=resnet_out_scale_factor,
+        )
+    elif down_block_type == "AttnDownBlock2D":
+        if add_downsample is False:
+            downsample_type = None
+        else:
+            downsample_type = downsample_type or "conv"  # default to 'conv'
+        return AttnDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            downsample_type=downsample_type,
+        )
+    elif down_block_type == "CrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlock2D")
+        return CrossAttnDownBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    # custom MV2D attention block
+    elif down_block_type == "CrossAttnDownBlockMV2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnDownBlockMV2D")
+        return CrossAttnDownBlockMV2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            num_views=num_views,
+            cd_attention_last=cd_attention_last,
+            cd_attention_mid=cd_attention_mid,
+            multiview_attention=multiview_attention,
+            sparse_mv_attention=sparse_mv_attention,
+            selfattn_block=selfattn_block,
+            mvcd_attention=mvcd_attention,
+            use_dino=use_dino
+        )
+    elif down_block_type == "SimpleCrossAttnDownBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnDownBlock2D")
+        return SimpleCrossAttnDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            skip_time_act=resnet_skip_time_act,
+            output_scale_factor=resnet_out_scale_factor,
+            only_cross_attention=only_cross_attention,
+            cross_attention_norm=cross_attention_norm,
+        )
+    elif down_block_type == "SkipDownBlock2D":
+        return SkipDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "AttnSkipDownBlock2D":
+        return AttnSkipDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "DownEncoderBlock2D":
+        return DownEncoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "AttnDownEncoderBlock2D":
+        return AttnDownEncoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            downsample_padding=downsample_padding,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif down_block_type == "KDownBlock2D":
+        return KDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+        )
+    elif down_block_type == "KCrossAttnDownBlock2D":
+        return KCrossAttnDownBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_downsample=add_downsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            add_self_attention=True if not add_downsample else False,
+        )
+    raise ValueError(f"{down_block_type} does not exist.")
+
+
+def get_up_block(
+    up_block_type,
+    num_layers,
+    in_channels,
+    out_channels,
+    prev_output_channel,
+    temb_channels,
+    add_upsample,
+    resnet_eps,
+    resnet_act_fn,
+    transformer_layers_per_block=1,
+    num_attention_heads=None,
+    resnet_groups=None,
+    cross_attention_dim=None,
+    dual_cross_attention=False,
+    use_linear_projection=False,
+    only_cross_attention=False,
+    upcast_attention=False,
+    resnet_time_scale_shift="default",
+    resnet_skip_time_act=False,
+    resnet_out_scale_factor=1.0,
+    cross_attention_norm=None,
+    attention_head_dim=None,
+    upsample_type=None,
+    num_views=1,
+    cd_attention_last: bool = False,
+    cd_attention_mid: bool = False,
+    multiview_attention: bool = True,
+    sparse_mv_attention: bool = False,
+    selfattn_block: str = "custom",
+    mvcd_attention: bool=False,
+    use_dino: bool = False  
+):
+    # If attn head dim is not defined, we default it to the number of heads
+    if attention_head_dim is None:
+        logger.warn(
+            f"It is recommended to provide `attention_head_dim` when calling `get_up_block`. Defaulting `attention_head_dim` to {num_attention_heads}."
+        )
+        attention_head_dim = num_attention_heads
+
+    up_block_type = up_block_type[7:] if up_block_type.startswith("UNetRes") else up_block_type
+    if up_block_type == "UpBlock2D":
+        return UpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "ResnetUpsampleBlock2D":
+        return ResnetUpsampleBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            skip_time_act=resnet_skip_time_act,
+            output_scale_factor=resnet_out_scale_factor,
+        )
+    elif up_block_type == "CrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlock2D")
+        return CrossAttnUpBlock2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    # custom MV2D attention block
+    elif up_block_type == "CrossAttnUpBlockMV2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for CrossAttnUpBlockMV2D")
+        return CrossAttnUpBlockMV2D(
+            num_layers=num_layers,
+            transformer_layers_per_block=transformer_layers_per_block,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            num_attention_heads=num_attention_heads,
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            only_cross_attention=only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            num_views=num_views,
+            cd_attention_last=cd_attention_last,
+            cd_attention_mid=cd_attention_mid,
+            multiview_attention=multiview_attention,
+            sparse_mv_attention=sparse_mv_attention,
+            selfattn_block=selfattn_block,
+            mvcd_attention=mvcd_attention,
+            use_dino=use_dino
+        )    
+    elif up_block_type == "SimpleCrossAttnUpBlock2D":
+        if cross_attention_dim is None:
+            raise ValueError("cross_attention_dim must be specified for SimpleCrossAttnUpBlock2D")
+        return SimpleCrossAttnUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            skip_time_act=resnet_skip_time_act,
+            output_scale_factor=resnet_out_scale_factor,
+            only_cross_attention=only_cross_attention,
+            cross_attention_norm=cross_attention_norm,
+        )
+    elif up_block_type == "AttnUpBlock2D":
+        if add_upsample is False:
+            upsample_type = None
+        else:
+            upsample_type = upsample_type or "conv"  # default to 'conv'
+
+        return AttnUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            upsample_type=upsample_type,
+        )
+    elif up_block_type == "SkipUpBlock2D":
+        return SkipUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "AttnSkipUpBlock2D":
+        return AttnSkipUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            prev_output_channel=prev_output_channel,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+        )
+    elif up_block_type == "UpDecoderBlock2D":
+        return UpDecoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            temb_channels=temb_channels,
+        )
+    elif up_block_type == "AttnUpDecoderBlock2D":
+        return AttnUpDecoderBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            resnet_groups=resnet_groups,
+            attention_head_dim=attention_head_dim,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            temb_channels=temb_channels,
+        )
+    elif up_block_type == "KUpBlock2D":
+        return KUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+        )
+    elif up_block_type == "KCrossAttnUpBlock2D":
+        return KCrossAttnUpBlock2D(
+            num_layers=num_layers,
+            in_channels=in_channels,
+            out_channels=out_channels,
+            temb_channels=temb_channels,
+            add_upsample=add_upsample,
+            resnet_eps=resnet_eps,
+            resnet_act_fn=resnet_act_fn,
+            cross_attention_dim=cross_attention_dim,
+            attention_head_dim=attention_head_dim,
+        )
+
+    raise ValueError(f"{up_block_type} does not exist.")
+
+
+class UNetMidBlockMV2DCrossAttn(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads=1,
+        output_scale_factor=1.0,
+        cross_attention_dim=1280,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        upcast_attention=False,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        sparse_mv_attention: bool = False,
+        selfattn_block: str = "custom", 
+        mvcd_attention: bool=False,
+        use_dino: bool = False
+    ):
+        super().__init__()
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        resnet_groups = resnet_groups if resnet_groups is not None else min(in_channels // 4, 32)
+        if selfattn_block == "custom":
+            from .transformer_mv2d import TransformerMV2DModel
+        elif selfattn_block == "rowwise":
+            from .transformer_mv2d_rowwise import TransformerMV2DModel
+        elif selfattn_block == "self_rowwise":
+            from .transformer_mv2d_self_rowwise import TransformerMV2DModel
+        else:
+            raise NotImplementedError
+        
+        # there is always at least one resnet
+        resnets = [
+            ResnetBlock2D(
+                in_channels=in_channels,
+                out_channels=in_channels,
+                temb_channels=temb_channels,
+                eps=resnet_eps,
+                groups=resnet_groups,
+                dropout=dropout,
+                time_embedding_norm=resnet_time_scale_shift,
+                non_linearity=resnet_act_fn,
+                output_scale_factor=output_scale_factor,
+                pre_norm=resnet_pre_norm,
+            )
+        ]
+        attentions = []
+
+        for _ in range(num_layers):
+            if not dual_cross_attention:
+                attentions.append(
+                    TransformerMV2DModel(
+                        num_attention_heads,
+                        in_channels // num_attention_heads,
+                        in_channels=in_channels,
+                        num_layers=transformer_layers_per_block,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        upcast_attention=upcast_attention,
+                        num_views=num_views,
+                        cd_attention_last=cd_attention_last,
+                        cd_attention_mid=cd_attention_mid,
+                        multiview_attention=multiview_attention,
+                        sparse_mv_attention=sparse_mv_attention,
+                        mvcd_attention=mvcd_attention,
+                        use_dino=use_dino
+                    )
+                )
+            else:
+                raise NotImplementedError
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=in_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        dino_feature: Optional[torch.FloatTensor] = None
+    ) -> torch.FloatTensor:
+        hidden_states = self.resnets[0](hidden_states, temb)
+        for attn, resnet in zip(self.attentions, self.resnets[1:]):
+            hidden_states = attn(
+                hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                cross_attention_kwargs=cross_attention_kwargs,
+                attention_mask=attention_mask,
+                encoder_attention_mask=encoder_attention_mask,
+                dino_feature=dino_feature,
+                return_dict=False,
+            )[0]
+            hidden_states = resnet(hidden_states, temb)
+
+        return hidden_states
+
+
+class CrossAttnUpBlockMV2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        prev_output_channel: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        add_upsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        sparse_mv_attention: bool = False,
+        selfattn_block: str = "custom",
+        mvcd_attention: bool=False,
+        use_dino: bool = False
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+
+        if selfattn_block == "custom":
+            from .transformer_mv2d import TransformerMV2DModel
+        elif selfattn_block == "rowwise":
+            from .transformer_mv2d_rowwise import TransformerMV2DModel
+        elif selfattn_block == "self_rowwise":
+            from .transformer_mv2d_self_rowwise import TransformerMV2DModel
+        else:
+            raise NotImplementedError
+        
+        for i in range(num_layers):
+            res_skip_channels = in_channels if (i == num_layers - 1) else out_channels
+            resnet_in_channels = prev_output_channel if i == 0 else out_channels
+
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=resnet_in_channels + res_skip_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    TransformerMV2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        num_views=num_views,
+                        cd_attention_last=cd_attention_last,
+                        cd_attention_mid=cd_attention_mid,
+                        multiview_attention=multiview_attention,
+                        sparse_mv_attention=sparse_mv_attention,
+                        mvcd_attention=mvcd_attention,
+                        use_dino=use_dino
+                    )
+                )
+            else:
+                raise NotImplementedError
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_upsample:
+            self.upsamplers = nn.ModuleList([Upsample2D(out_channels, use_conv=True, out_channels=out_channels)])
+        else:
+            self.upsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        res_hidden_states_tuple: Tuple[torch.FloatTensor, ...],
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        upsample_size: Optional[int] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        dino_feature: Optional[torch.FloatTensor] = None
+    ):
+        for resnet, attn in zip(self.resnets, self.attentions):
+            # pop res hidden states
+            res_hidden_states = res_hidden_states_tuple[-1]
+            res_hidden_states_tuple = res_hidden_states_tuple[:-1]
+            hidden_states = torch.cat([hidden_states, res_hidden_states], dim=1)
+
+            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(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    dino_feature,
+                    None,  # timestep
+                    None,  # class_labels
+                    cross_attention_kwargs,
+                    attention_mask,
+                    encoder_attention_mask,
+                    **ckpt_kwargs,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    dino_feature=dino_feature,
+                    return_dict=False,
+                )[0]
+
+        if self.upsamplers is not None:
+            for upsampler in self.upsamplers:
+                hidden_states = upsampler(hidden_states, upsample_size)
+
+        return hidden_states
+
+
+class CrossAttnDownBlockMV2D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        temb_channels: int,
+        dropout: float = 0.0,
+        num_layers: int = 1,
+        transformer_layers_per_block: int = 1,
+        resnet_eps: float = 1e-6,
+        resnet_time_scale_shift: str = "default",
+        resnet_act_fn: str = "swish",
+        resnet_groups: int = 32,
+        resnet_pre_norm: bool = True,
+        num_attention_heads=1,
+        cross_attention_dim=1280,
+        output_scale_factor=1.0,
+        downsample_padding=1,
+        add_downsample=True,
+        dual_cross_attention=False,
+        use_linear_projection=False,
+        only_cross_attention=False,
+        upcast_attention=False,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        sparse_mv_attention: bool = False,
+        selfattn_block: str = "custom",
+        mvcd_attention: bool=False,
+        use_dino: bool = False
+    ):
+        super().__init__()
+        resnets = []
+        attentions = []
+
+        self.has_cross_attention = True
+        self.num_attention_heads = num_attention_heads
+        if selfattn_block == "custom":
+            from .transformer_mv2d import TransformerMV2DModel
+        elif selfattn_block == "rowwise":
+            from .transformer_mv2d_rowwise import TransformerMV2DModel
+        elif selfattn_block == "self_rowwise":
+            from .transformer_mv2d_self_rowwise import TransformerMV2DModel
+        else:
+            raise NotImplementedError
+        
+        for i in range(num_layers):
+            in_channels = in_channels if i == 0 else out_channels
+            resnets.append(
+                ResnetBlock2D(
+                    in_channels=in_channels,
+                    out_channels=out_channels,
+                    temb_channels=temb_channels,
+                    eps=resnet_eps,
+                    groups=resnet_groups,
+                    dropout=dropout,
+                    time_embedding_norm=resnet_time_scale_shift,
+                    non_linearity=resnet_act_fn,
+                    output_scale_factor=output_scale_factor,
+                    pre_norm=resnet_pre_norm,
+                )
+            )
+            if not dual_cross_attention:
+                attentions.append(
+                    TransformerMV2DModel(
+                        num_attention_heads,
+                        out_channels // num_attention_heads,
+                        in_channels=out_channels,
+                        num_layers=transformer_layers_per_block,
+                        cross_attention_dim=cross_attention_dim,
+                        norm_num_groups=resnet_groups,
+                        use_linear_projection=use_linear_projection,
+                        only_cross_attention=only_cross_attention,
+                        upcast_attention=upcast_attention,
+                        num_views=num_views,
+                        cd_attention_last=cd_attention_last,
+                        cd_attention_mid=cd_attention_mid,
+                        multiview_attention=multiview_attention,
+                        sparse_mv_attention=sparse_mv_attention,
+                        mvcd_attention=mvcd_attention,
+                        use_dino=use_dino
+                    )
+                )
+            else:
+                raise NotImplementedError
+        self.attentions = nn.ModuleList(attentions)
+        self.resnets = nn.ModuleList(resnets)
+
+        if add_downsample:
+            self.downsamplers = nn.ModuleList(
+                [
+                    Downsample2D(
+                        out_channels, use_conv=True, out_channels=out_channels, padding=downsample_padding, name="op"
+                    )
+                ]
+            )
+        else:
+            self.downsamplers = None
+
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        temb: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        dino_feature: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        additional_residuals=None,
+    ):
+        output_states = ()
+
+        blocks = list(zip(self.resnets, self.attentions))
+
+        for i, (resnet, attn) in enumerate(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(resnet),
+                    hidden_states,
+                    temb,
+                    **ckpt_kwargs,
+                )
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(attn, return_dict=False),
+                    hidden_states,
+                    encoder_hidden_states,
+                    dino_feature,
+                    None,  # timestep
+                    None,  # class_labels
+                    cross_attention_kwargs,
+                    attention_mask,
+                    encoder_attention_mask,
+                    **ckpt_kwargs,
+                )[0]
+            else:
+                hidden_states = resnet(hidden_states, temb)
+                hidden_states = attn(
+                    hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    dino_feature=dino_feature,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    return_dict=False,
+                )[0]
+
+            # apply additional residuals to the output of the last pair of resnet and attention blocks
+            if i == len(blocks) - 1 and additional_residuals is not None:
+                hidden_states = hidden_states + additional_residuals
+
+            output_states = output_states + (hidden_states,)
+
+        if self.downsamplers is not None:
+            for downsampler in self.downsamplers:
+                hidden_states = downsampler(hidden_states)
+
+            output_states = output_states + (hidden_states,)
+
+        return hidden_states, output_states
+

mvdiffusion/models/unet_mv2d_condition.py

@@ -0,0 +1,1686 @@
+# 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.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+import os
+
+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 BaseOutput, logging
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import AttentionProcessor, AttnProcessor
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin, load_state_dict, _load_state_dict_into_model
+from diffusers.models.unet_2d_blocks import (
+    CrossAttnDownBlock2D,
+    CrossAttnUpBlock2D,
+    DownBlock2D,
+    UNetMidBlock2DCrossAttn,
+    UNetMidBlock2DSimpleCrossAttn,
+    UpBlock2D,
+)
+from diffusers.utils import (
+    CONFIG_NAME,
+    FLAX_WEIGHTS_NAME,
+    SAFETENSORS_WEIGHTS_NAME,
+    WEIGHTS_NAME,
+    _add_variant,
+    _get_model_file,
+    deprecate,
+    is_torch_version,
+    logging,
+)
+from diffusers.utils.import_utils import is_accelerate_available 
+from diffusers.utils.hub_utils import HF_HUB_OFFLINE
+from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
+DIFFUSERS_CACHE = HUGGINGFACE_HUB_CACHE
+
+from diffusers import __version__
+from .unet_mv2d_blocks import (
+    CrossAttnDownBlockMV2D,
+    CrossAttnUpBlockMV2D,
+    UNetMidBlockMV2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+from einops import rearrange, repeat
+
+from diffusers import __version__
+from mvdiffusion.models.unet_mv2d_blocks import (
+    CrossAttnDownBlockMV2D,
+    CrossAttnUpBlockMV2D,
+    UNetMidBlockMV2DCrossAttn,
+    get_down_block,
+    get_up_block,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNetMV2DConditionOutput(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 ResidualBlock(nn.Module):
+    def __init__(self, dim):
+        super(ResidualBlock, self).__init__()
+        self.linear1 = nn.Linear(dim, dim)
+        self.activation = nn.SiLU()
+        self.linear2 = nn.Linear(dim, dim)
+
+    def forward(self, x):
+        identity = x
+        out = self.linear1(x)
+        out = self.activation(out)
+        out = self.linear2(out)
+        out += identity  
+        out = self.activation(out)
+        return out
+    
+class ResidualLiner(nn.Module):
+    def __init__(self, in_features, out_features, dim, act=None, num_block=1):
+        super(ResidualLiner, self).__init__()
+        self.linear_in = nn.Sequential(nn.Linear(in_features, dim), nn.SiLU())
+        
+        blocks = nn.ModuleList()
+        for _ in range(num_block):
+            blocks.append(ResidualBlock(dim))
+        self.blocks = blocks    
+        
+        self.linear_out = nn.Linear(dim, out_features)
+        self.act = act
+
+    def forward(self, x):
+        out = self.linear_in(x)
+        for block in self.blocks:
+            out = block(out)
+        out = self.linear_out(out)
+        if self.act is not None:
+            out = self.act(out)
+        return out
+
+class BasicConvBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, stride=1):
+        super(BasicConvBlock, self).__init__()
+        self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.norm1 = nn.GroupNorm(num_groups=8, num_channels=in_channels,  affine=True)
+        self.act = nn.SiLU()
+        self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)
+        self.norm2 = nn.GroupNorm(num_groups=8, num_channels=in_channels,  affine=True)
+        self.downsample = nn.Sequential()
+        if stride != 1 or in_channels != out_channels:
+            self.downsample = nn.Sequential(
+                nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=stride, bias=False),
+                nn.GroupNorm(num_groups=8, num_channels=in_channels,  affine=True)
+            )
+
+    def forward(self, x):
+        identity = x
+        out = self.conv1(x)
+        out = self.norm1(out)
+        out = self.act(out)
+        out = self.conv2(out)
+        out = self.norm2(out)
+        out += self.downsample(identity)
+        out = self.act(out)
+        return out
+
+class UNetMV2DConditionModel(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 either `UNetMidBlock2DCrossAttn` 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.
+        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` or `Tuple[int]`, *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`].
+        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] = (
+            "CrossAttnDownBlockMV2D",
+            "CrossAttnDownBlockMV2D",
+            "CrossAttnDownBlockMV2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlockMV2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlockMV2D", "CrossAttnUpBlockMV2D", "CrossAttnUpBlockMV2D"),
+        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,
+        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]] = 1,
+        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,
+        projection_camera_embeddings_input_dim: Optional[int] = None,
+        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,
+        num_views: int = 1,
+        cd_attention_last: bool = False,
+        cd_attention_mid: bool = False,
+        multiview_attention: bool = True,
+        sparse_mv_attention: bool = False,
+        selfattn_block: str = "custom",
+        mvcd_attention: bool = False,
+        regress_elevation: bool = False,
+        regress_focal_length: bool = False,
+        num_regress_blocks: int = 4,
+        use_dino: bool = False,
+        addition_downsample: bool = False,
+        addition_channels: Optional[Tuple[int]] = (1280, 1280, 1280),
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+        self.num_views = num_views
+        self.mvcd_attention = mvcd_attention
+        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}."
+            )
+
+        # 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,
+                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,
+                num_views=num_views,
+                cd_attention_last=cd_attention_last,
+                cd_attention_mid=cd_attention_mid,
+                multiview_attention=multiview_attention,
+                sparse_mv_attention=sparse_mv_attention,
+                selfattn_block=selfattn_block,
+                mvcd_attention=mvcd_attention,
+                use_dino=use_dino
+            )
+            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,
+                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,
+            )
+        # custom MV2D attention block  
+        elif mid_block_type == "UNetMidBlockMV2DCrossAttn":
+            self.mid_block = UNetMidBlockMV2DCrossAttn(
+                transformer_layers_per_block=transformer_layers_per_block[-1],
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                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,
+                num_views=num_views,
+                cd_attention_last=cd_attention_last,
+                cd_attention_mid=cd_attention_mid,
+                multiview_attention=multiview_attention,
+                sparse_mv_attention=sparse_mv_attention,
+                selfattn_block=selfattn_block,
+                mvcd_attention=mvcd_attention,
+                use_dino=use_dino
+            )
+        elif mid_block_type == "UNetMidBlock2DSimpleCrossAttn":
+            self.mid_block = UNetMidBlock2DSimpleCrossAttn(
+                in_channels=block_out_channels[-1],
+                temb_channels=blocks_time_embed_dim,
+                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 is None:
+            self.mid_block = None
+        else:
+            raise ValueError(f"unknown mid_block_type : {mid_block_type}")
+        
+        self.addition_downsample = addition_downsample
+        if self.addition_downsample:
+            inc = block_out_channels[-1]
+            self.downsample = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
+            self.conv_block = nn.ModuleList()
+            self.conv_block.append(BasicConvBlock(inc, addition_channels[0], stride=1)) 
+            for dim_ in addition_channels[1:-1]:  
+                self.conv_block.append(BasicConvBlock(dim_, dim_, stride=1))    
+            self.conv_block.append(BasicConvBlock(dim_, inc))
+            self.addition_conv_out = nn.Conv2d(inc, inc, kernel_size=1, bias=False)
+            nn.init.zeros_(self.addition_conv_out.weight.data)
+            self.addition_act_out = nn.SiLU()
+            self.upsample = nn.Upsample(scale_factor=2, mode='bilinear', align_corners=True)
+                
+        self.regress_elevation = regress_elevation
+        self.regress_focal_length = regress_focal_length
+        if regress_elevation or regress_focal_length:
+            self.pool = nn.AdaptiveAvgPool2d((1, 1))
+            self.camera_embedding = TimestepEmbedding(projection_camera_embeddings_input_dim, time_embed_dim=time_embed_dim) 
+        
+        regress_in_dim = block_out_channels[-1]*2 if mvcd_attention else block_out_channels
+            
+        if regress_elevation:
+            self.elevation_regressor = ResidualLiner(regress_in_dim, 1, 1280, act=None, num_block=num_regress_blocks)
+        if regress_focal_length:
+            self.focal_regressor = ResidualLiner(regress_in_dim, 1, 1280, act=None, num_block=num_regress_blocks)
+        '''
+        self.regress_elevation = regress_elevation
+        self.regress_focal_length = regress_focal_length
+        if regress_elevation and (not regress_focal_length):
+            print("Regressing elevation")
+            cam_dim = 1
+        elif regress_focal_length and (not regress_elevation):
+            print("Regressing focal length")
+            cam_dim = 6
+        elif regress_elevation and regress_focal_length:
+            print("Regressing both elevation and focal length")
+            cam_dim = 7
+        else:
+            cam_dim = 0
+        assert projection_camera_embeddings_input_dim == 2*cam_dim, "projection_camera_embeddings_input_dim should be 2*cam_dim"
+        if regress_elevation or regress_focal_length:
+            self.elevation_regressor = nn.ModuleList([
+                nn.Linear(block_out_channels[-1], 1280),
+                nn.SiLU(),
+                nn.Linear(1280, 1280),
+                nn.SiLU(),
+                nn.Linear(1280, cam_dim)
+            ])
+            self.pool = nn.AdaptiveAvgPool2d((1, 1))
+            self.focal_act = nn.Softmax(dim=-1)
+            self.camera_embedding = TimestepEmbedding(projection_camera_embeddings_input_dim, time_embed_dim=time_embed_dim) 
+        '''
+          
+        # 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))
+        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,
+                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,
+                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,
+                num_views=num_views,
+                cd_attention_last=cd_attention_last,
+                cd_attention_mid=cd_attention_mid,
+                multiview_attention=multiview_attention,
+                sparse_mv_attention=sparse_mv_attention,
+                selfattn_block=selfattn_block,
+                mvcd_attention=mvcd_attention,
+                use_dino=use_dino
+            )
+            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
+        )
+
+    @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, "set_processor"):
+                processors[f"{name}.processor"] = module.processor
+
+            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]]):
+        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)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            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.
+        """
+        self.set_attn_processor(AttnProcessor())
+
+    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 isinstance(module, (CrossAttnDownBlock2D, CrossAttnDownBlockMV2D, DownBlock2D, CrossAttnUpBlock2D, CrossAttnUpBlockMV2D, UpBlock2D)):
+            module.gradient_checkpointing = value
+
+    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,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        dino_feature: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        vis_max_min: bool = False,
+    ) -> Union[UNetMV2DConditionOutput, 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)`.
+            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.
+
+        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.
+        """
+        record_max_min = {}
+        # 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
+
+        if any(s % default_overall_up_factor != 0 for s in sample.shape[-2:]):
+            logger.info("Forward upsample size to force interpolation output size.")
+            forward_upsample_size = True
+
+        # 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
+        emb_pre_act = 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)
+        # 2. pre-process
+        sample = self.conv_in(sample)
+        # 3. down
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        is_adapter = mid_block_additional_residual is None and down_block_additional_residuals is not None
+
+        down_block_res_samples = (sample,)
+        for i, downsample_block in enumerate(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_block_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_block_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    dino_feature=dino_feature,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+
+                if is_adapter and len(down_block_additional_residuals) > 0:
+                    sample += down_block_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
+
+        if self.addition_downsample:
+            global_sample = sample
+            global_sample = self.downsample(global_sample)
+            for layer in self.conv_block:
+                global_sample = layer(global_sample)
+            global_sample = self.addition_act_out(self.addition_conv_out(global_sample))
+            global_sample = self.upsample(global_sample)
+        # 4. mid
+        if self.mid_block is not None:
+            sample = self.mid_block(
+                sample,
+                emb,
+                encoder_hidden_states=encoder_hidden_states,
+                dino_feature=dino_feature,
+                attention_mask=attention_mask,
+                cross_attention_kwargs=cross_attention_kwargs,
+                encoder_attention_mask=encoder_attention_mask,
+            )        
+        # 4.1 regress elevation and focal length
+        # # predict elevation -> embed -> projection -> add to time emb
+        if self.regress_elevation or self.regress_focal_length:
+            pool_embeds = self.pool(sample.detach()).squeeze(-1).squeeze(-1) # (2B, C)
+            if self.mvcd_attention:
+                pool_embeds_normal, pool_embeds_color = torch.chunk(pool_embeds, 2, dim=0)
+                pool_embeds = torch.cat([pool_embeds_normal, pool_embeds_color], dim=-1) # (B, 2C)
+            pose_pred = []
+            if self.regress_elevation:
+                ele_pred = self.elevation_regressor(pool_embeds)
+                ele_pred = rearrange(ele_pred, '(b v) c -> b v c', v=self.num_views)
+                ele_pred = torch.mean(ele_pred, dim=1)
+                pose_pred.append(ele_pred) # b, c
+            
+            if self.regress_focal_length:
+                focal_pred = self.focal_regressor(pool_embeds)
+                focal_pred = rearrange(focal_pred, '(b v) c -> b v c', v=self.num_views)
+                focal_pred = torch.mean(focal_pred, dim=1)
+                pose_pred.append(focal_pred)
+            pose_pred = torch.cat(pose_pred, dim=-1)
+            # 'e_de_da_sincos', (B, 2)
+            pose_embeds = torch.cat([
+                torch.sin(pose_pred),
+                torch.cos(pose_pred)
+            ], dim=-1)
+            pose_embeds = self.camera_embedding(pose_embeds)
+            pose_embeds = torch.repeat_interleave(pose_embeds, self.num_views, 0) 
+            if self.mvcd_attention:
+                pose_embeds = torch.cat([pose_embeds,] * 2, dim=0)
+
+            emb = pose_embeds + emb_pre_act
+            if self.time_embed_act is not None:
+                emb = self.time_embed_act(emb)
+            
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        if self.addition_downsample:
+            sample = sample + global_sample
+            
+        # 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,
+                    dino_feature=dino_feature,
+                    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
+                )
+        if torch.isnan(sample).any() or torch.isinf(sample).any():
+            print("NAN in sample, stop training.")
+            exit() 
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+        if not return_dict:
+            return (sample, pose_pred)
+        if self.regress_elevation or self.regress_focal_length:
+            return UNetMV2DConditionOutput(sample=sample), pose_pred
+        else:
+            return UNetMV2DConditionOutput(sample=sample)
+
+        
+    @classmethod
+    def from_pretrained_2d(
+            cls, pretrained_model_name_or_path: Optional[Union[str, os.PathLike]],
+            camera_embedding_type: str, num_views: int, sample_size: int,
+            zero_init_conv_in: bool = True, zero_init_camera_projection: bool = False,
+            projection_camera_embeddings_input_dim: int=2,  
+            cd_attention_last: bool = False, num_regress_blocks: int = 4,
+            cd_attention_mid: bool = False, multiview_attention: bool = True, 
+            sparse_mv_attention: bool = False, selfattn_block: str = 'custom', mvcd_attention: bool = False,
+            in_channels: int = 8, out_channels: int = 4, unclip: bool = False, regress_elevation: bool = False, regress_focal_length: bool = False, 
+            init_mvattn_with_selfattn: bool= False, use_dino: bool = False, addition_downsample: bool = False,
+            **kwargs
+        ):
+        r"""
+        Instantiate a pretrained PyTorch model from a pretrained model configuration.
+
+        The model is set in evaluation mode - `model.eval()` - by default, and dropout modules are deactivated. To
+        train the model, set it back in training mode with `model.train()`.
+
+        Parameters:
+            pretrained_model_name_or_path (`str` or `os.PathLike`, *optional*):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`~ModelMixin.save_pretrained`].
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            torch_dtype (`str` or `torch.dtype`, *optional*):
+                Override the default `torch.dtype` and load the model with another dtype. If `"auto"` is passed, the
+                dtype is automatically derived from the model's weights.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+            resume_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to resume downloading the model weights and configuration files. If set to `False`, any
+                incompletely downloaded files are deleted.
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            output_loading_info (`bool`, *optional*, defaults to `False`):
+                Whether or not to also return a dictionary containing missing keys, unexpected keys and error messages.
+            local_files_only(`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            use_auth_token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            from_flax (`bool`, *optional*, defaults to `False`):
+                Load the model weights from a Flax checkpoint save file.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+            mirror (`str`, *optional*):
+                Mirror source to resolve accessibility issues if you're downloading a model in China. We do not
+                guarantee the timeliness or safety of the source, and you should refer to the mirror site for more
+                information.
+            device_map (`str` or `Dict[str, Union[int, str, torch.device]]`, *optional*):
+                A map that specifies where each submodule should go. It doesn't need to be defined for each
+                parameter/buffer name; once a given module name is inside, every submodule of it will be sent to the
+                same device.
+
+                Set `device_map="auto"` to have 🤗 Accelerate automatically compute the most optimized `device_map`. For
+                more information about each option see [designing a device
+                map](https://hf.co/docs/accelerate/main/en/usage_guides/big_modeling#designing-a-device-map).
+            max_memory (`Dict`, *optional*):
+                A dictionary device identifier for the maximum memory. Will default to the maximum memory available for
+                each GPU and the available CPU RAM if unset.
+            offload_folder (`str` or `os.PathLike`, *optional*):
+                The path to offload weights if `device_map` contains the value `"disk"`.
+            offload_state_dict (`bool`, *optional*):
+                If `True`, temporarily offloads the CPU state dict to the hard drive to avoid running out of CPU RAM if
+                the weight of the CPU state dict + the biggest shard of the checkpoint does not fit. Defaults to `True`
+                when there is some disk offload.
+            low_cpu_mem_usage (`bool`, *optional*, defaults to `True` if torch version >= 1.9.0 else `False`):
+                Speed up model loading only loading the pretrained weights and not initializing the weights. This also
+                tries to not use more than 1x model size in CPU memory (including peak memory) while loading the model.
+                Only supported for PyTorch >= 1.9.0. If you are using an older version of PyTorch, setting this
+                argument to `True` will raise an error.
+            variant (`str`, *optional*):
+                Load weights from a specified `variant` filename such as `"fp16"` or `"ema"`. This is ignored when
+                loading `from_flax`.
+            use_safetensors (`bool`, *optional*, defaults to `None`):
+                If set to `None`, the `safetensors` weights are downloaded if they're available **and** if the
+                `safetensors` library is installed. If set to `True`, the model is forcibly loaded from `safetensors`
+                weights. If set to `False`, `safetensors` weights are not loaded.
+
+        <Tip>
+
+        To use private or [gated models](https://huggingface.co/docs/hub/models-gated#gated-models), log-in with
+        `huggingface-cli login`. You can also activate the special
+        ["offline-mode"](https://huggingface.co/diffusers/installation.html#offline-mode) to use this method in a
+        firewalled environment.
+
+        </Tip>
+
+        Example:
+
+        ```py
+        from diffusers import UNet2DConditionModel
+
+        unet = UNet2DConditionModel.from_pretrained("runwayml/stable-diffusion-v1-5", subfolder="unet")
+        ```
+
+        If you get the error message below, you need to finetune the weights for your downstream task:
+
+        ```bash
+        Some weights of UNet2DConditionModel were not initialized from the model checkpoint at runwayml/stable-diffusion-v1-5 and are newly initialized because the shapes did not match:
+        - conv_in.weight: found shape torch.Size([320, 4, 3, 3]) in the checkpoint and torch.Size([320, 9, 3, 3]) in the model instantiated
+        You should probably TRAIN this model on a down-stream task to be able to use it for predictions and inference.
+        ```
+        """
+        cache_dir = kwargs.pop("cache_dir", DIFFUSERS_CACHE)
+        ignore_mismatched_sizes = kwargs.pop("ignore_mismatched_sizes", False)
+        force_download = kwargs.pop("force_download", False)
+        from_flax = kwargs.pop("from_flax", False)
+        resume_download = kwargs.pop("resume_download", False)
+        proxies = kwargs.pop("proxies", None)
+        output_loading_info = kwargs.pop("output_loading_info", False)
+        local_files_only = kwargs.pop("local_files_only", HF_HUB_OFFLINE)
+        use_auth_token = kwargs.pop("use_auth_token", None)
+        revision = kwargs.pop("revision", None)
+        torch_dtype = kwargs.pop("torch_dtype", None)
+        subfolder = kwargs.pop("subfolder", None)
+        device_map = kwargs.pop("device_map", None)
+        max_memory = kwargs.pop("max_memory", None)
+        offload_folder = kwargs.pop("offload_folder", None)
+        offload_state_dict = kwargs.pop("offload_state_dict", False)
+        variant = kwargs.pop("variant", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        if use_safetensors:
+            raise ValueError(
+                "`use_safetensors`=True but safetensors is not installed. Please install safetensors with `pip install safetensors"
+            )
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        if device_map is not None and not is_accelerate_available():
+            raise NotImplementedError(
+                "Loading and dispatching requires `accelerate`. Please make sure to install accelerate or set"
+                " `device_map=None`. You can install accelerate with `pip install accelerate`."
+            )
+
+        # Check if we can handle device_map and dispatching the weights
+        if device_map is not None and not is_torch_version(">=", "1.9.0"):
+            raise NotImplementedError(
+                "Loading and dispatching requires torch >= 1.9.0. Please either update your PyTorch version or set"
+                " `device_map=None`."
+            )
+
+        # Load config if we don't provide a configuration
+        config_path = pretrained_model_name_or_path
+
+        user_agent = {
+            "diffusers": __version__,
+            "file_type": "model",
+            "framework": "pytorch",
+        }
+
+        # load config
+        config, unused_kwargs, commit_hash = cls.load_config(
+            config_path,
+            cache_dir=cache_dir,
+            return_unused_kwargs=True,
+            return_commit_hash=True,
+            force_download=force_download,
+            resume_download=resume_download,
+            proxies=proxies,
+            local_files_only=local_files_only,
+            use_auth_token=use_auth_token,
+            revision=revision,
+            subfolder=subfolder,
+            device_map=device_map,
+            max_memory=max_memory,
+            offload_folder=offload_folder,
+            offload_state_dict=offload_state_dict,
+            user_agent=user_agent,
+            **kwargs,
+        )
+
+        # modify config
+        config["_class_name"] = cls.__name__
+        config['in_channels'] = in_channels
+        config['out_channels'] = out_channels
+        config['sample_size'] = sample_size # training resolution
+        config['num_views'] = num_views
+        config['cd_attention_last'] = cd_attention_last
+        config['cd_attention_mid'] = cd_attention_mid
+        config['multiview_attention'] = multiview_attention
+        config['sparse_mv_attention'] = sparse_mv_attention
+        config['selfattn_block'] = selfattn_block
+        config['mvcd_attention'] = mvcd_attention
+        config["down_block_types"] = [
+            "CrossAttnDownBlockMV2D",
+            "CrossAttnDownBlockMV2D",
+            "CrossAttnDownBlockMV2D",
+            "DownBlock2D"
+        ]
+        config['mid_block_type'] = "UNetMidBlockMV2DCrossAttn"
+        config["up_block_types"] = [
+            "UpBlock2D",
+            "CrossAttnUpBlockMV2D",
+            "CrossAttnUpBlockMV2D",
+            "CrossAttnUpBlockMV2D"
+        ]
+        
+
+        config['regress_elevation'] = regress_elevation # true
+        config['regress_focal_length'] = regress_focal_length # true
+        config['projection_camera_embeddings_input_dim'] = projection_camera_embeddings_input_dim # 2 for elevation and 10 for focal_length  
+        config['use_dino'] = use_dino 
+        config['num_regress_blocks'] = num_regress_blocks
+        config['addition_downsample'] = addition_downsample
+        # load model
+        model_file = None
+        if from_flax:
+            raise NotImplementedError
+        else:
+            if use_safetensors:
+                try:
+                    model_file = _get_model_file(
+                        pretrained_model_name_or_path,
+                        weights_name=_add_variant(SAFETENSORS_WEIGHTS_NAME, variant),
+                        cache_dir=cache_dir,
+                        force_download=force_download,
+                        resume_download=resume_download,
+                        proxies=proxies,
+                        local_files_only=local_files_only,
+                        use_auth_token=use_auth_token,
+                        revision=revision,
+                        subfolder=subfolder,
+                        user_agent=user_agent,
+                        commit_hash=commit_hash,
+                    )
+                except IOError as e:
+                    if not allow_pickle:
+                        raise e
+                    pass
+            if model_file is None:
+                model_file = _get_model_file(
+                    pretrained_model_name_or_path,
+                    weights_name=_add_variant(WEIGHTS_NAME, variant),
+                    cache_dir=cache_dir,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    user_agent=user_agent,
+                    commit_hash=commit_hash,
+                )
+
+            model = cls.from_config(config, **unused_kwargs)
+            import copy
+            state_dict_pretrain = load_state_dict(model_file, variant=variant)
+            state_dict = copy.deepcopy(state_dict_pretrain)
+            
+            if init_mvattn_with_selfattn:
+                for key in state_dict_pretrain:
+                    if 'attn1' in key:
+                        key_mv = key.replace('attn1', 'attn_mv')
+                        state_dict[key_mv] = state_dict_pretrain[key]
+                        if 'to_out.0.weight' in key:
+                            nn.init.zeros_(state_dict[key_mv].data)
+                    if 'transformer_blocks' in key and 'norm1' in key: # in case that initialize the norm layer in resnet block
+                        key_mv = key.replace('norm1', 'norm_mv')
+                        state_dict[key_mv] = state_dict_pretrain[key]
+            # del state_dict_pretrain
+            
+            model._convert_deprecated_attention_blocks(state_dict)
+
+            conv_in_weight = state_dict['conv_in.weight']
+            conv_out_weight = state_dict['conv_out.weight']
+            model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model_2d(
+                model,
+                state_dict,
+                model_file,
+                pretrained_model_name_or_path,
+                ignore_mismatched_sizes=True,
+            )
+            if any([key == 'conv_in.weight' for key, _, _ in mismatched_keys]):
+                # initialize from the original SD structure
+                model.conv_in.weight.data[:,:4] = conv_in_weight
+
+            # whether to place all zero to new layers?
+            if zero_init_conv_in:
+                model.conv_in.weight.data[:,4:] = 0.
+
+            if any([key == 'conv_out.weight' for key, _, _ in mismatched_keys]):
+                # initialize from the original SD structure
+                model.conv_out.weight.data[:,:4] = conv_out_weight
+                if out_channels == 8: # copy for the last 4 channels
+                    model.conv_out.weight.data[:, 4:] = conv_out_weight
+
+            if zero_init_camera_projection: # true
+                params = [p for p in model.camera_embedding.parameters()]
+                torch.nn.init.zeros_(params[-1].data)
+
+            loading_info = {
+                "missing_keys": missing_keys,
+                "unexpected_keys": unexpected_keys,
+                "mismatched_keys": mismatched_keys,
+                "error_msgs": error_msgs,
+            }
+
+        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
+            raise ValueError(
+                f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
+            )
+        elif torch_dtype is not None:
+            model = model.to(torch_dtype)
+
+        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+
+        # Set model in evaluation mode to deactivate DropOut modules by default
+        model.eval()
+        if output_loading_info:
+            return model, loading_info
+        return model
+
+    @classmethod
+    def _load_pretrained_model_2d(
+        cls,
+        model,
+        state_dict,
+        resolved_archive_file,
+        pretrained_model_name_or_path,
+        ignore_mismatched_sizes=False,
+    ):
+        # Retrieve missing & unexpected_keys
+        model_state_dict = model.state_dict()
+        loaded_keys = list(state_dict.keys())
+
+        expected_keys = list(model_state_dict.keys())
+
+        original_loaded_keys = loaded_keys
+
+        missing_keys = list(set(expected_keys) - set(loaded_keys))
+        unexpected_keys = list(set(loaded_keys) - set(expected_keys))
+
+        # Make sure we are able to load base models as well as derived models (with heads)
+        model_to_load = model
+
+        def _find_mismatched_keys(
+            state_dict,
+            model_state_dict,
+            loaded_keys,
+            ignore_mismatched_sizes,
+        ):
+            mismatched_keys = []
+            if ignore_mismatched_sizes:
+                for checkpoint_key in loaded_keys:
+                    model_key = checkpoint_key
+
+                    if (
+                        model_key in model_state_dict
+                        and state_dict[checkpoint_key].shape != model_state_dict[model_key].shape
+                    ):
+                        mismatched_keys.append(
+                            (checkpoint_key, state_dict[checkpoint_key].shape, model_state_dict[model_key].shape)
+                        )
+                        del state_dict[checkpoint_key]
+            return mismatched_keys
+
+        if state_dict is not None:
+            # Whole checkpoint
+            mismatched_keys = _find_mismatched_keys(
+                state_dict,
+                model_state_dict,
+                original_loaded_keys,
+                ignore_mismatched_sizes,
+            )
+            error_msgs = _load_state_dict_into_model(model_to_load, state_dict)
+
+        if len(error_msgs) > 0:
+            error_msg = "\n\t".join(error_msgs)
+            if "size mismatch" in error_msg:
+                error_msg += (
+                    "\n\tYou may consider adding `ignore_mismatched_sizes=True` in the model `from_pretrained` method."
+                )
+            raise RuntimeError(f"Error(s) in loading state_dict for {model.__class__.__name__}:\n\t{error_msg}")
+
+        if len(unexpected_keys) > 0:
+            logger.warning(
+                f"Some weights of the model checkpoint at {pretrained_model_name_or_path} were not used when"
+                f" initializing {model.__class__.__name__}: {unexpected_keys}\n- This IS expected if you are"
+                f" initializing {model.__class__.__name__} from the checkpoint of a model trained on another task"
+                " or with another architecture (e.g. initializing a BertForSequenceClassification model from a"
+                " BertForPreTraining model).\n- This IS NOT expected if you are initializing"
+                f" {model.__class__.__name__} from the checkpoint of a model that you expect to be exactly"
+                " identical (initializing a BertForSequenceClassification model from a"
+                " BertForSequenceClassification model)."
+            )
+        else:
+            logger.info(f"All model checkpoint weights were used when initializing {model.__class__.__name__}.\n")
+        if len(missing_keys) > 0:
+            logger.warning(
+                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+                f" {pretrained_model_name_or_path} and are newly initialized: {missing_keys}\nYou should probably"
+                " TRAIN this model on a down-stream task to be able to use it for predictions and inference."
+            )
+        elif len(mismatched_keys) == 0:
+            logger.info(
+                f"All the weights of {model.__class__.__name__} were initialized from the model checkpoint at"
+                f" {pretrained_model_name_or_path}.\nIf your task is similar to the task the model of the"
+                f" checkpoint was trained on, you can already use {model.__class__.__name__} for predictions"
+                " without further training."
+            )
+        if len(mismatched_keys) > 0:
+            mismatched_warning = "\n".join(
+                [
+                    f"- {key}: found shape {shape1} in the checkpoint and {shape2} in the model instantiated"
+                    for key, shape1, shape2 in mismatched_keys
+                ]
+            )
+            logger.warning(
+                f"Some weights of {model.__class__.__name__} were not initialized from the model checkpoint at"
+                f" {pretrained_model_name_or_path} and are newly initialized because the shapes did not"
+                f" match:\n{mismatched_warning}\nYou should probably TRAIN this model on a down-stream task to be"
+                " able to use it for predictions and inference."
+            )
+
+        return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs
+

mvdiffusion/pipelines/pipeline_mvdiffusion_unclip.py

@@ -0,0 +1,633 @@
+import inspect
+import warnings
+from typing import Callable, List, Optional, Union, Dict, Any
+import PIL
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection, CLIPFeatureExtractor, CLIPTokenizer, CLIPTextModel
+from diffusers.utils.import_utils import is_accelerate_available
+from diffusers.configuration_utils import FrozenDict
+from diffusers.image_processor import VaeImageProcessor
+from diffusers.models import AutoencoderKL, UNet2DConditionModel
+from diffusers.models.embeddings import get_timestep_embedding
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import deprecate, logging
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
+from diffusers.pipelines.stable_diffusion.stable_unclip_image_normalizer import StableUnCLIPImageNormalizer
+import os
+import torchvision.transforms.functional as TF
+from einops import rearrange
+logger = logging.get_logger(__name__)
+
+class StableUnCLIPImg2ImgPipeline(DiffusionPipeline):
+    """
+    Pipeline for text-guided image to image generation using stable unCLIP.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods the
+    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)
+
+    Args:
+        feature_extractor ([`CLIPFeatureExtractor`]):
+            Feature extractor for image pre-processing before being encoded.
+        image_encoder ([`CLIPVisionModelWithProjection`]):
+            CLIP vision model for encoding images.
+        image_normalizer ([`StableUnCLIPImageNormalizer`]):
+            Used to normalize the predicted image embeddings before the noise is applied and un-normalize the image
+            embeddings after the noise has been applied.
+        image_noising_scheduler ([`KarrasDiffusionSchedulers`]):
+            Noise schedule for adding noise to the predicted image embeddings. The amount of noise to add is determined
+            by `noise_level` in `StableUnCLIPPipeline.__call__`.
+        tokenizer (`CLIPTokenizer`):
+            Tokenizer of class
+            [CLIPTokenizer](https://huggingface.co/docs/transformers/v4.21.0/en/model_doc/clip#transformers.CLIPTokenizer).
+        text_encoder ([`CLIPTextModel`]):
+            Frozen text-encoder.
+        unet ([`UNet2DConditionModel`]): Conditional U-Net architecture to denoise the encoded image latents.
+        scheduler ([`KarrasDiffusionSchedulers`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents.
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
+    """
+    # image encoding components
+    feature_extractor: CLIPFeatureExtractor
+    image_encoder: CLIPVisionModelWithProjection
+    # image noising components
+    image_normalizer: StableUnCLIPImageNormalizer
+    image_noising_scheduler: KarrasDiffusionSchedulers
+    # regular denoising components
+    tokenizer: CLIPTokenizer
+    text_encoder: CLIPTextModel
+    unet: UNet2DConditionModel
+    scheduler: KarrasDiffusionSchedulers
+    vae: AutoencoderKL
+
+    def __init__(
+        self,
+        # image encoding components
+        feature_extractor: CLIPFeatureExtractor,
+        image_encoder: CLIPVisionModelWithProjection,
+        # image noising components
+        image_normalizer: StableUnCLIPImageNormalizer,
+        image_noising_scheduler: KarrasDiffusionSchedulers,
+        # regular denoising components
+        tokenizer: CLIPTokenizer,
+        text_encoder: CLIPTextModel,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        # vae
+        vae: AutoencoderKL,
+        num_views: int = 4,
+    ):
+        super().__init__()
+
+        self.register_modules(
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+            image_normalizer=image_normalizer,
+            image_noising_scheduler=image_noising_scheduler,
+            tokenizer=tokenizer,
+            text_encoder=text_encoder,
+            unet=unet,
+            scheduler=scheduler,
+            vae=vae,
+        )
+        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        self.num_views: int = num_views
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.enable_vae_slicing
+    def enable_vae_slicing(self):
+        r"""
+        Enable sliced VAE decoding.
+
+        When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several
+        steps. This is useful to save some memory and allow larger batch sizes.
+        """
+        self.vae.enable_slicing()
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.disable_vae_slicing
+    def disable_vae_slicing(self):
+        r"""
+        Disable sliced VAE decoding. If `enable_vae_slicing` was previously invoked, this method will go back to
+        computing decoding in one step.
+        """
+        self.vae.disable_slicing()
+
+    def enable_sequential_cpu_offload(self, gpu_id=0):
+        r"""
+        Offloads all models to CPU using accelerate, significantly reducing memory usage. When called, the pipeline's
+        models have their state dicts saved to CPU and then are moved to a `torch.device('meta') and loaded to GPU only
+        when their specific submodule has its `forward` method called.
+        """
+        if is_accelerate_available():
+            from accelerate import cpu_offload
+        else:
+            raise ImportError("Please install accelerate via `pip install accelerate`")
+
+        device = torch.device(f"cuda:{gpu_id}")
+
+        # TODO: self.image_normalizer.{scale,unscale} are not covered by the offload hooks, so they fails if added to the list
+        models = [
+            self.image_encoder,
+            self.text_encoder,
+            self.unet,
+            self.vae,
+        ]
+        for cpu_offloaded_model in models:
+            if cpu_offloaded_model is not None:
+                cpu_offload(cpu_offloaded_model, device)
+
+    @property
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._execution_device
+    def _execution_device(self):
+        r"""
+        Returns the device on which the pipeline's models will be executed. After calling
+        `pipeline.enable_sequential_cpu_offload()` the execution device can only be inferred from Accelerate's module
+        hooks.
+        """
+        if not hasattr(self.unet, "_hf_hook"):
+            return self.device
+        for module in self.unet.modules():
+            if (
+                hasattr(module, "_hf_hook")
+                and hasattr(module._hf_hook, "execution_device")
+                and module._hf_hook.execution_device is not None
+            ):
+                return torch.device(module._hf_hook.execution_device)
+        return self.device
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline._encode_prompt
+    def _encode_prompt(
+        self,
+        prompt,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        negative_prompt=None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        lora_scale: Optional[float] = None,
+    ):
+        r"""
+        Encodes the prompt into text encoder hidden states.
+
+        Args:
+             prompt (`str` or `List[str]`, *optional*):
+                prompt to be encoded
+            device: (`torch.device`):
+                torch device
+            num_images_per_prompt (`int`):
+                number of images that should be generated per prompt
+            do_classifier_free_guidance (`bool`):
+                whether to use classifier free guidance or not
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+        """
+        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)
+
+        if do_classifier_free_guidance:
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            normal_prompt_embeds, color_prompt_embeds = torch.chunk(prompt_embeds, 2, dim=0)
+            
+            prompt_embeds = torch.cat([normal_prompt_embeds, normal_prompt_embeds, color_prompt_embeds, color_prompt_embeds], 0)
+
+        return prompt_embeds
+
+    def _encode_image(
+        self,
+        image_pil,
+        device,
+        num_images_per_prompt,
+        do_classifier_free_guidance,
+        noise_level: int=0,
+        generator: Optional[torch.Generator] = None
+    ):
+        dtype = next(self.image_encoder.parameters()).dtype
+        # ______________________________clip image embedding______________________________ 
+        image = self.feature_extractor(images=image_pil, return_tensors="pt").pixel_values
+        image = image.to(device=device, dtype=dtype)
+        image_embeds = self.image_encoder(image).image_embeds
+        
+        image_embeds = self.noise_image_embeddings(
+            image_embeds=image_embeds,
+            noise_level=noise_level,
+            generator=generator,
+            )
+        # duplicate image embeddings for each generation per prompt, using mps friendly method
+        # image_embeds = image_embeds.unsqueeze(1)
+        # note: the condition input is same
+        image_embeds = image_embeds.repeat(num_images_per_prompt, 1)
+
+        if do_classifier_free_guidance:
+            normal_image_embeds, color_image_embeds = torch.chunk(image_embeds, 2, dim=0)
+            negative_prompt_embeds = torch.zeros_like(normal_image_embeds)
+
+            # For classifier free guidance, we need to do two forward passes.
+            # Here we concatenate the unconditional and text embeddings into a single batch
+            # to avoid doing two forward passes
+            image_embeds = torch.cat([negative_prompt_embeds, normal_image_embeds, negative_prompt_embeds, color_image_embeds], 0)
+            
+        # _____________________________vae input latents__________________________________________________
+        image_pt = torch.stack([TF.to_tensor(img) for img in image_pil], dim=0).to(device)
+        image_pt = image_pt * 2.0 - 1.0
+        image_latents = self.vae.encode(image_pt).latent_dist.mode() * self.vae.config.scaling_factor
+        # Note: repeat differently from official pipelines     
+        image_latents = image_latents.repeat(num_images_per_prompt, 1, 1, 1)
+
+        if do_classifier_free_guidance:
+            normal_image_latents, color_image_latents = torch.chunk(image_latents, 2, dim=0)
+            image_latents = torch.cat([torch.zeros_like(normal_image_latents), normal_image_latents, 
+                                       torch.zeros_like(color_image_latents), color_image_latents], 0)
+
+        return image_embeds, image_latents
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.decode_latents
+    def decode_latents(self, latents):
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents).sample
+        image = (image / 2 + 0.5).clamp(0, 1)
+        # we always cast to float32 as this does not cause significant overhead and is compatible with bfloat16
+        image = image.cpu().permute(0, 2, 3, 1).float().numpy()
+        return image
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
+    def prepare_extra_step_kwargs(self, generator, eta):
+        # prepare extra kwargs for the scheduler step, since not all schedulers have the same signature
+        # eta (η) is only used with the DDIMScheduler, it will be ignored for other schedulers.
+        # eta corresponds to η in DDIM paper: https://arxiv.org/abs/2010.02502
+        # and should be between [0, 1]
+
+        accepts_eta = "eta" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        extra_step_kwargs = {}
+        if accepts_eta:
+            extra_step_kwargs["eta"] = eta
+
+        # check if the scheduler accepts generator
+        accepts_generator = "generator" in set(inspect.signature(self.scheduler.step).parameters.keys())
+        if accepts_generator:
+            extra_step_kwargs["generator"] = generator
+        return extra_step_kwargs
+
+    def check_inputs(
+        self,
+        prompt,
+        image,
+        height,
+        width,
+        callback_steps,
+        noise_level,
+    ):
+        if height % 8 != 0 or width % 8 != 0:
+            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+
+        if (callback_steps is None) or (
+            callback_steps is not None and (not isinstance(callback_steps, int) or callback_steps <= 0)
+        ):
+            raise ValueError(
+                f"`callback_steps` has to be a positive integer but is {callback_steps} of type"
+                f" {type(callback_steps)}."
+            )
+
+        if prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
+            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
+
+
+        if noise_level < 0 or noise_level >= self.image_noising_scheduler.config.num_train_timesteps:
+            raise ValueError(
+                f"`noise_level` must be between 0 and {self.image_noising_scheduler.config.num_train_timesteps - 1}, inclusive."
+            )
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_latents
+    def prepare_latents(self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None):
+        shape = (batch_size, num_channels_latents, height // self.vae_scale_factor, width // self.vae_scale_factor)
+        if isinstance(generator, list) and len(generator) != batch_size:
+            raise ValueError(
+                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
+                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
+            )
+
+        if latents is None:
+            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
+        else:
+            latents = latents.to(device)
+
+        # scale the initial noise by the standard deviation required by the scheduler
+        latents = latents * self.scheduler.init_noise_sigma
+        return latents
+
+    # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_unclip.StableUnCLIPPipeline.noise_image_embeddings
+    def noise_image_embeddings(
+        self,
+        image_embeds: torch.Tensor,
+        noise_level: int,
+        noise: Optional[torch.FloatTensor] = None,
+        generator: Optional[torch.Generator] = None,
+    ):
+        """
+        Add noise to the image embeddings. The amount of noise is controlled by a `noise_level` input. A higher
+        `noise_level` increases the variance in the final un-noised images.
+
+        The noise is applied in two ways
+        1. A noise schedule is applied directly to the embeddings
+        2. A vector of sinusoidal time embeddings are appended to the output.
+
+        In both cases, the amount of noise is controlled by the same `noise_level`.
+
+        The embeddings are normalized before the noise is applied and un-normalized after the noise is applied.
+        """
+        if noise is None:
+            noise = randn_tensor(
+                image_embeds.shape, generator=generator, device=image_embeds.device, dtype=image_embeds.dtype
+            )
+
+        noise_level = torch.tensor([noise_level] * image_embeds.shape[0], device=image_embeds.device)
+
+        image_embeds = self.image_normalizer.scale(image_embeds)
+
+        image_embeds = self.image_noising_scheduler.add_noise(image_embeds, timesteps=noise_level, noise=noise)
+
+        image_embeds = self.image_normalizer.unscale(image_embeds)
+
+        noise_level = get_timestep_embedding(
+            timesteps=noise_level, embedding_dim=image_embeds.shape[-1], flip_sin_to_cos=True, downscale_freq_shift=0
+        )
+
+        # `get_timestep_embeddings` does not contain any weights and will always return f32 tensors,
+        # but we might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        noise_level = noise_level.to(image_embeds.dtype)
+
+        image_embeds = torch.cat((image_embeds, noise_level), 1)
+
+        return image_embeds
+
+    @torch.no_grad()
+    # @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        image: Union[torch.FloatTensor, PIL.Image.Image],
+        prompt: Union[str, List[str]],   
+        prompt_embeds: torch.FloatTensor = None,
+        dino_feature: torch.FloatTensor = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 20,
+        guidance_scale: float = 10,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[torch.Generator] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        callback: Optional[Callable[[int, int, torch.FloatTensor], None]] = None,
+        callback_steps: int = 1,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        noise_level: int = 0,
+        image_embeds: Optional[torch.FloatTensor] = None,
+        return_elevation_focal: Optional[bool] = False,
+        gt_img_in: Optional[torch.FloatTensor] = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            image (`torch.FloatTensor` or `PIL.Image.Image`):
+                `Image`, or tensor representing an image batch. The image will be encoded to its CLIP embedding which
+                the unet will be conditioned on. Note that the image is _not_ encoded by the vae and then used as the
+                latents in the denoising process such as in the standard stable diffusion text guided image variation
+                process.
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image.
+            num_inference_steps (`int`, *optional*, defaults to 20):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            guidance_scale (`float`, *optional*, defaults to 10.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts not to guide the image generation. If not defined, one has to pass
+                `negative_prompt_embeds`. instead. If not defined, one has to pass `negative_prompt_embeds`. instead.
+                Ignored when not using guidance (i.e., ignored if `guidance_scale` is less than `1`).
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            eta (`float`, *optional*, defaults to 0.0):
+                Corresponds to parameter eta (η) in the DDIM paper: https://arxiv.org/abs/2010.02502. Only applies to
+                [`schedulers.DDIMScheduler`], will be ignored for others.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
+                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
+                argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] instead of a
+                plain tuple.
+            callback (`Callable`, *optional*):
+                A function that will be called every `callback_steps` steps during inference. The function will be
+                called with the following arguments: `callback(step: int, timestep: int, latents: torch.FloatTensor)`.
+            callback_steps (`int`, *optional*, defaults to 1):
+                The frequency at which the `callback` function will be called. If not specified, the callback will be
+                called at every step.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttnProcessor` as defined under
+                `self.processor` in
+                [diffusers.cross_attention](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/cross_attention.py).
+            noise_level (`int`, *optional*, defaults to `0`):
+                The amount of noise to add to the image embeddings. A higher `noise_level` increases the variance in
+                the final un-noised images. See `StableUnCLIPPipeline.noise_image_embeddings` for details.
+            image_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated CLIP embeddings to condition the unet on. Note that these are not latents to be used in
+                the denoising process. If you want to provide pre-generated latents, pass them to `__call__` as
+                `latents`.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.ImagePipelineOutput`] or `tuple`: [`~ pipeline_utils.ImagePipelineOutput`] if `return_dict` is
+            True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated images.
+        """
+        # 0. Default height and width to unet
+        height = height or self.unet.config.sample_size * self.vae_scale_factor
+        width = width or self.unet.config.sample_size * self.vae_scale_factor
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt=prompt,
+            image=image,
+            height=height,
+            width=width,
+            callback_steps=callback_steps,
+            noise_level=noise_level
+        )
+
+        # 2. Define call parameters
+        if isinstance(image, list):
+            batch_size = len(image)
+        elif isinstance(image, torch.Tensor):
+            batch_size = image.shape[0]
+            assert batch_size >= self.num_views and batch_size % self.num_views == 0
+        elif isinstance(image, PIL.Image.Image):
+            image = [image]*self.num_views*2
+            batch_size = self.num_views*2
+
+        if isinstance(prompt, str):
+            prompt = [prompt] * self.num_views * 2
+
+        device = self._execution_device
+
+        # here `guidance_scale` is defined analog to the guidance weight `w` of equation (2)
+        # of the Imagen paper: https://arxiv.org/pdf/2205.11487.pdf . `guidance_scale = 1`
+        # corresponds to doing no classifier free guidance.
+        do_classifier_free_guidance = guidance_scale != 1.0
+
+        # 3. Encode input prompt
+        text_encoder_lora_scale = (
+            cross_attention_kwargs.get("scale", None) if cross_attention_kwargs is not None else None
+        )
+        prompt_embeds = self._encode_prompt(
+            prompt=prompt,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            negative_prompt=negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=text_encoder_lora_scale,
+        )
+        
+
+        # 4. Encoder input image
+        if isinstance(image, list):
+            image_pil = image
+        elif isinstance(image, torch.Tensor):
+            image_pil = [TF.to_pil_image(image[i]) for i in range(image.shape[0])]
+        noise_level = torch.tensor([noise_level], device=device)
+        image_embeds, image_latents = self._encode_image(
+            image_pil=image_pil,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            do_classifier_free_guidance=do_classifier_free_guidance,
+            noise_level=noise_level,
+            generator=generator,
+        )
+
+        # 5. Prepare timesteps
+        self.scheduler.set_timesteps(num_inference_steps, device=device)
+        timesteps = self.scheduler.timesteps
+
+        # 6. Prepare latent variables
+        num_channels_latents = self.unet.config.out_channels
+        if gt_img_in is not None:
+            latents = gt_img_in * self.scheduler.init_noise_sigma
+        else:
+            latents = self.prepare_latents(
+                batch_size=batch_size,
+                num_channels_latents=num_channels_latents,
+                height=height,
+                width=width,
+                dtype=prompt_embeds.dtype,
+                device=device,
+                generator=generator,
+                latents=latents,
+            )
+
+        # 7. Prepare extra step kwargs. TODO: Logic should ideally just be moved out of the pipeline
+        extra_step_kwargs = self.prepare_extra_step_kwargs(generator, eta)
+
+        eles, focals = [], []
+        # 8. Denoising loop
+        for i, t in enumerate(self.progress_bar(timesteps)):
+            if do_classifier_free_guidance:
+                normal_latents, color_latents = torch.chunk(latents, 2, dim=0)  
+                latent_model_input = torch.cat([normal_latents, normal_latents, color_latents, color_latents], 0)
+            else:
+                latent_model_input = latents
+            latent_model_input = torch.cat([
+                    latent_model_input, image_latents
+                ], dim=1)
+            latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+            # predict the noise residual
+            unet_out = self.unet(
+                latent_model_input,
+                t,
+                encoder_hidden_states=prompt_embeds,
+                dino_feature=dino_feature,
+                class_labels=image_embeds,
+                cross_attention_kwargs=cross_attention_kwargs,
+                return_dict=False)
+            
+            noise_pred = unet_out[0]
+            if return_elevation_focal:    
+                uncond_pose, pose  = torch.chunk(unet_out[1], 2, 0) 
+                pose = uncond_pose + guidance_scale * (pose - uncond_pose)
+                ele = pose[:, 0].detach().cpu().numpy() # b
+                eles.append(ele)
+                focal = pose[:, 1].detach().cpu().numpy()
+                focals.append(focal)
+                
+            # perform guidance
+            if do_classifier_free_guidance:
+                normal_noise_pred_uncond, normal_noise_pred_text, color_noise_pred_uncond, color_noise_pred_text = torch.chunk(noise_pred, 4, dim=0)
+                
+                noise_pred_uncond, noise_pred_text = torch.cat([normal_noise_pred_uncond, color_noise_pred_uncond], 0), torch.cat([normal_noise_pred_text, color_noise_pred_text], 0)
+                noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+            # compute the previous noisy sample x_t -> x_t-1
+            latents = self.scheduler.step(noise_pred, t, latents, **extra_step_kwargs, return_dict=False)[0]
+
+            if callback is not None and i % callback_steps == 0:
+                callback(i, t, latents)
+
+        # 9. Post-processing
+        if not output_type == "latent":
+            if num_channels_latents == 8:
+                latents = torch.cat([latents[:, :4], latents[:, 4:]], dim=0)
+            with torch.no_grad():
+                image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False)[0]
+        else:
+            image = latents
+
+        image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload last model to CPU
+        # if hasattr(self, "final_offload_hook") and self.final_offload_hook is not None:
+        #     self.final_offload_hook.offload()
+        if not return_dict:
+            return (image, )
+        if return_elevation_focal:
+            return ImagePipelineOutput(images=image),  eles, focals
+        else:
+            return ImagePipelineOutput(images=image)