svd/temporal_controlnet.py

# 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.
'''
    This is a ControlNet for sptio temporal unet (SVD)
'''
from dataclasses import dataclass
from typing import Any, Dict, List, Optional, Tuple, Union
import os, sys
import random
import torch
from torch import nn
from torch.nn import functional as F

from diffusers import AutoencoderKLTemporalDecoder
from diffusers.configuration_utils import ConfigMixin, register_to_config
from diffusers.loaders import FromOriginalControlnetMixin
from diffusers.utils import BaseOutput, logging
from diffusers.models.attention_processor import (
    ADDED_KV_ATTENTION_PROCESSORS,
    CROSS_ATTENTION_PROCESSORS,
    AttentionProcessor,
    AttnAddedKVProcessor,
    AttnProcessor,
)
from diffusers.models.embeddings import TimestepEmbedding, Timesteps
from diffusers.models.modeling_utils import ModelMixin

# Import files from the local folder
root_path = os.path.abspath('.')
sys.path.append(root_path)
from svd.unet_spatio_temporal_condition import UNetSpatioTemporalConditionModel
from svd.diffusion_arch.unet_3d_blocks import UNetMidBlockSpatioTemporal, get_down_block

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


def zero_module(module):
    for p in module.parameters():
        nn.init.zeros_(p)
    return module


@dataclass
class ControlNetOutput(BaseOutput):
    """
    The output of [`ControlNetModel`].

    Args:
        down_block_res_samples (`tuple[torch.Tensor]`):
            A tuple of downsample activations at different resolutions for each downsampling block. Each tensor should
            be of shape `(batch_size, channel * resolution, height //resolution, width // resolution)`. Output can be
            used to condition the original UNet's downsampling activations.
        mid_down_block_re_sample (`torch.Tensor`):
            The activation of the midde block (the lowest sample resolution). Each tensor should be of shape
            `(batch_size, channel * lowest_resolution, height // lowest_resolution, width // lowest_resolution)`.
            Output can be used to condition the original UNet's middle block activation.
    """

    down_block_res_samples: Tuple[torch.Tensor]
    mid_block_res_sample: torch.Tensor



class ControlNetModel(ModelMixin, ConfigMixin, FromOriginalControlnetMixin):
    """
    A ControlNet model.

    Args:
        in_channels (`int`, defaults to 4):
            The number of channels in the input sample.
        flip_sin_to_cos (`bool`, defaults to `True`):
            Whether to flip the sin to cos in the time embedding.
        freq_shift (`int`, defaults to 0):
            The frequency shift to apply to the time embedding.
        down_block_types (`tuple[str]`, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
            The tuple of downsample blocks to use.
        only_cross_attention (`Union[bool, Tuple[bool]]`, defaults to `False`):
        block_out_channels (`tuple[int]`, defaults to `(320, 640, 1280, 1280)`):
            The tuple of output channels for each block.
        layers_per_block (`int`, defaults to 2):
            The number of layers per block.
        downsample_padding (`int`, defaults to 1):
            The padding to use for the downsampling convolution.
        mid_block_scale_factor (`float`, defaults to 1):
            The scale factor to use for the mid block.
        act_fn (`str`, 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`, defaults to 1e-5):
            The epsilon to use for the normalization.
        cross_attention_dim (`int`, 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 (`Union[int, Tuple[int]]`, defaults to 8):
            The dimension of the attention heads.
        use_linear_projection (`bool`, defaults to `False`):
        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.
        num_class_embeds (`int`, *optional*, defaults to 0):
            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`.
        upcast_attention (`bool`, defaults to `False`):
        resnet_time_scale_shift (`str`, defaults to `"default"`):
            Time scale shift config for ResNet blocks (see `ResnetBlock2D`). Choose from `default` or `scale_shift`.
        projection_class_embeddings_input_dim (`int`, *optional*, defaults to `None`):
            The dimension of the `class_labels` input when `class_embed_type="projection"`. Required when
            `class_embed_type="projection"`.
        controlnet_conditioning_channel_order (`str`, defaults to `"rgb"`):
            The channel order of conditional image. Will convert to `rgb` if it's `bgr`.
        conditioning_embedding_out_channels (`tuple[int]`, *optional*, defaults to `(16, 32, 96, 256)`):
            The tuple of output channel for each block in the `conditioning_embedding` layer.
        global_pool_conditions (`bool`, defaults to `False`):
            TODO(Patrick) - unused parameter.
        addition_embed_type_num_heads (`int`, defaults to 64):
            The number of heads to use for the `TextTimeEmbedding` layer.
    """

    _supports_gradient_checkpointing = True

    @register_to_config
    def __init__(
        self,
        in_channels: int = 8,
        conditioning_channels: int = 3,
        flip_sin_to_cos: bool = True,
        freq_shift: int = 0,
        down_block_types: Tuple[str, ...] = (
            "CrossAttnDownBlockSpatioTemporal",
            "CrossAttnDownBlockSpatioTemporal",
            "CrossAttnDownBlockSpatioTemporal",
            "DownBlockSpatioTemporal",
        ),
        mid_block_type: Optional[str] = "UNetMidBlockSpatioTemporal",
        block_out_channels: Tuple[int, ...] = (320, 640, 1280, 1280),
        addition_time_embed_dim: int = 256,
        layers_per_block: int = 2,
        act_fn: str = "silu",
        cross_attention_dim: int = 1024,     
        projection_class_embeddings_input_dim: Optional[int] = 768,
        conditioning_embedding_out_channels: Optional[Tuple[int, ...]] = (16, 32, 96, 256),
        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
        num_attention_heads: Union[int, Tuple[int]] = (5, 10, 20, 20),    # This is modified to SVD config setting for the default case
        encoder_hid_dim: Optional[int] = None,
        encoder_hid_dim_type: Optional[str] = None,
        controlnet_conditioning_channel_order = 'rgb',
    ):
        super().__init__()

        self.controlnet_conditioning_channel_order = controlnet_conditioning_channel_order

        # Check inputs
        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(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 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}."
            )

        

        ########################## First convolution channel for sample (noise) ##########################
        conv_in_kernel = 3
        conv_in_padding = (conv_in_kernel - 1) // 2
        self.conv_in_concat = zero_module(nn.Conv2d(
            12, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
        ))      # Input is 12 channels (8 + 4) right now
        

        ########################## Time embedding and so on ##########################
        time_embed_dim = block_out_channels[0] * 4

        self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)  # defualt flip_sin_to_cos True
        timestep_input_dim = block_out_channels[0]
        self.time_embedding = TimestepEmbedding(
            timestep_input_dim,
            time_embed_dim,
            act_fn=act_fn,
        )       

        # Additional time embedding for other purpose
        self.add_time_proj = Timesteps(addition_time_embed_dim, True, downscale_freq_shift=0)       # This will include hyperparameter like fps, motion_bucket_id, noise_aug_strength
        self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)

        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}."
            )



        #############################  Down and Mid Blocks Init  #############################
        # Init ModuleList and prepare information needed
        self.down_blocks = nn.ModuleList([])
        output_channel = block_out_channels[0]


        # Check instance
        if isinstance(num_attention_heads, int):
            num_attention_heads = (num_attention_heads,) * 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)

        blocks_time_embed_dim = time_embed_dim
        

        # ControlNet Module!!!!!
        self.controlnet_down_blocks = nn.ModuleList([])     
        controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
        controlnet_block = zero_module(controlnet_block)        # Zero Convolution
        self.controlnet_down_blocks.append(controlnet_block)    

        # Down block init one by one
        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=1e-5,
                cross_attention_dim=cross_attention_dim[i],
                num_attention_heads=num_attention_heads[i],
                resnet_act_fn="silu",
            )
            self.down_blocks.append(down_block)


            # ControlNet Module !!!!
            for _ in range(layers_per_block[0]):       # Loop 2 times here
                controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
                controlnet_block = zero_module(controlnet_block)         
                self.controlnet_down_blocks.append(controlnet_block)

            if not is_final_block:      # Loop only once
                controlnet_block = nn.Conv2d(output_channel, output_channel, kernel_size=1)
                controlnet_block = zero_module(controlnet_block)
                self.controlnet_down_blocks.append(controlnet_block)

        # Mid block 
        mid_block_channel = block_out_channels[-1]

        # ControlNet Module !!!!
        controlnet_block = nn.Conv2d(mid_block_channel, mid_block_channel, kernel_size=1)
        controlnet_block = zero_module(controlnet_block)
        self.controlnet_mid_block = controlnet_block

        if mid_block_type == "UNetMidBlockSpatioTemporal":     
            self.mid_block = UNetMidBlockSpatioTemporal(
                block_out_channels[-1],
                temb_channels=blocks_time_embed_dim,
                transformer_layers_per_block=transformer_layers_per_block[-1],
                cross_attention_dim=cross_attention_dim[-1],
                num_attention_heads=num_attention_heads[-1],
            )
        else:
            raise ValueError(f"unknown mid_block_type : {mid_block_type}")


    @classmethod
    def from_unet(
        cls,
        unet: UNetSpatioTemporalConditionModel,
        conditioning_channels: int = 3,
        load_weights_from_unet: bool = True,    
    ):
        r"""
        Instantiate a [`ControlNetModel`] from [`UNetSpatioTemporalConditionModel`].

        Parameters:
            unet (`UNetSpatioTemporalConditionModel`):
                The UNet model weights to copy to the [`ControlNetModel`]. All configuration options are also copied
                where applicable.
            load_weights_from_unet (bool):
                Whether we used unet as trainable copy (Should be True in default)
        """

        controlnet = cls(conditioning_channels=conditioning_channels)

        if load_weights_from_unet:
            # controlnet.conv_in.load_state_dict(unet.conv_in.state_dict())     # Won't load this conv_in now, we will replace it with another zero conv
            controlnet.time_proj.load_state_dict(unet.time_proj.state_dict())
            controlnet.time_embedding.load_state_dict(unet.time_embedding.state_dict())

            controlnet.down_blocks.load_state_dict(unet.down_blocks.state_dict())
            controlnet.mid_block.load_state_dict(unet.mid_block.state_dict())

        return controlnet
    

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

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

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

            return processors

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

        return processors


    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
        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.
        """
        if all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
            processor = AttnProcessor()
        else:
            raise ValueError(
                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
            )

        self.set_attn_processor(processor)


    def _set_gradient_checkpointing(self, module, value: bool = False) -> None:
        if hasattr(module, "gradient_checkpointing"):
            module.gradient_checkpointing = value


    def enable_forward_chunking(self, chunk_size: Optional[int] = None, dim: int = 0) -> None:
        """
        Sets the attention processor to use [feed forward
        chunking](https://huggingface.co/blog/reformer#2-chunked-feed-forward-layers).

        Parameters:
            chunk_size (`int`, *optional*):
                The chunk size of the feed-forward layers. If not specified, will run feed-forward layer individually
                over each tensor of dim=`dim`.
            dim (`int`, *optional*, defaults to `0`):
                The dimension over which the feed-forward computation should be chunked. Choose between dim=0 (batch)
                or dim=1 (sequence length).
        """
        if dim not in [0, 1]:
            raise ValueError(f"Make sure to set `dim` to either 0 or 1, not {dim}")

        # By default chunk size is 1
        chunk_size = chunk_size or 1

        def fn_recursive_feed_forward(module: torch.nn.Module, chunk_size: int, dim: int):
            if hasattr(module, "set_chunk_feed_forward"):
                module.set_chunk_feed_forward(chunk_size=chunk_size, dim=dim)

            for child in module.children():
                fn_recursive_feed_forward(child, chunk_size, dim)

        for module in self.children():
            fn_recursive_feed_forward(module, chunk_size, dim)


    def forward(
        self,
        sample: torch.FloatTensor,
        timestep: Union[torch.Tensor, float, int],
        encoder_hidden_states: torch.Tensor,
        added_time_ids: torch.Tensor,
        added_positions: torch.Tensor = None,
        controlnet_cond: torch.FloatTensor = None,
        conditioning_scale: float = 1.0,
        inner_conditioning_scale: float = 1.0,
        timestep_cond: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        guess_mode: bool = False,
        return_dict: bool = True,
    ) -> Union[ControlNetOutput, Tuple[Tuple[torch.FloatTensor, ...], torch.FloatTensor]]:
        """
        The [`ControlNetModel`] forward method.

        Args:
            sample (`torch.FloatTensor`):
                The noisy input tensor.
            timestep (`Union[torch.Tensor, float, int]`):
                The number of timesteps to denoise an input.
            encoder_hidden_states (`torch.Tensor`):
                The encoder hidden states.
            controlnet_cond (`torch.FloatTensor`):
                The conditional input tensor of shape `(batch_size, sequence_length, 4, hidden_size)` which is already encoded in VAE.
            conditioning_scale (`float`, defaults to `1.0`):
                The scale factor for ControlNet outputs.
            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
            timestep_cond (`torch.Tensor`, *optional*, defaults to `None`):
                Additional conditional embeddings for timestep. If provided, the embeddings will be summed with the
                timestep_embedding passed through the `self.time_embedding` layer to obtain the final timestep
                embeddings.
            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
                negative values to the attention scores corresponding to "discard" tokens.
            added_cond_kwargs (`dict`):
                Additional conditions for the Stable Diffusion XL UNet.
            cross_attention_kwargs (`dict[str]`, *optional*, defaults to `None`):
                A kwargs dictionary that if specified is passed along to the `AttnProcessor`.
            guess_mode (`bool`, defaults to `False`):
                In this mode, the ControlNet encoder tries its best to recognize the input content of the input even if
                you remove all prompts. A `guidance_scale` between 3.0 and 5.0 is recommended.
            return_dict (`bool`, defaults to `True`):
                Whether or not to return a [`~models.controlnet.ControlNetOutput`] instead of a plain tuple.

        Returns:
            [`~models.controlnet.ControlNetOutput`] **or** `tuple`:
                If `return_dict` is `True`, a [`~models.controlnet.ControlNetOutput`] is returned, otherwise a tuple is
                returned where the first element is the sample tensor.
        """ 

        # check channel order
        channel_order = self.controlnet_conditioning_channel_order

        # if channel_order == "rgb":
        #     # in rgb order by default
        #     ...
        # elif channel_order == "bgr":
        #     controlnet_cond = torch.flip(controlnet_cond, dims=[1])
        # else:
        #     raise ValueError(f"unknown `controlnet_conditioning_channel_order`: {channel_order}")

        # prepare attention_mask
        if attention_mask is not None:
            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
            attention_mask = attention_mask.unsqueeze(1)

        # 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
        batch_size, num_frames = sample.shape[:2]       # Take the classifier guidance also as an input in batch
        timesteps = timesteps.expand(batch_size)

        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)     # No more timestep_cond because usually this is None

        # motion score + fps + aug strength embeds
        time_embeds = self.add_time_proj(added_time_ids.flatten())
        time_embeds = time_embeds.reshape((batch_size, -1))
        time_embeds = time_embeds.to(emb.dtype)
        aug_emb = self.add_embedding(time_embeds)

        # Wrap up
        emb = emb + aug_emb


        sample = sample.flatten(0, 1)
        # Repeat the embeddings num_video_frames times
        # emb: [batch, channels] -> [batch * frames, channels]
        emb = emb.repeat_interleave(num_frames, dim=0)
        # encoder_hidden_states: [batch, 1, channels] -> [batch * frames, 1, channels]
        encoder_hidden_states = encoder_hidden_states.repeat_interleave(num_frames, dim=0)


        # 2. Pre-Process    
        image_only_indicator = torch.zeros(batch_size, num_frames, dtype=sample.dtype, device=sample.device)       


        # Feature: Concat the sample && controlnet_cond at dim 1 (channel-wise) !!!
        sample = torch.cat([sample, controlnet_cond], dim=1)


        # Merge sample and controlnet_cond together
        sample = self.conv_in_concat(sample)
        


        # 3. Down block
        down_block_res_samples = (sample,)
        for downsample_block in self.down_blocks:
            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
                sample, res_samples = downsample_block(
                    hidden_states=sample,   # Vae encode + noise
                    temb=emb,
                    encoder_hidden_states=encoder_hidden_states,        # Clip encode
                    image_only_indicator=image_only_indicator,
                )
            else:
                sample, res_samples = downsample_block(
                    hidden_states=sample,
                    temb=emb,
                    image_only_indicator=image_only_indicator,
                )

            down_block_res_samples += res_samples


        # 4. Mid block
        sample = self.mid_block(
            hidden_states=sample,
            temb=emb,
            encoder_hidden_states=encoder_hidden_states,
            image_only_indicator=image_only_indicator,
        )


        # 5. ControlNet blocks
        controlnet_down_block_res_samples = ()

        for down_block_res_sample, controlnet_block in zip(down_block_res_samples, self.controlnet_down_blocks):
            down_block_res_sample = controlnet_block(down_block_res_sample)
            controlnet_down_block_res_samples = controlnet_down_block_res_samples + (down_block_res_sample,)
        
        down_block_res_samples = controlnet_down_block_res_samples
        # Mid block
        mid_block_res_sample = self.controlnet_mid_block(sample)


        # 6. Scaling
        if guess_mode:
            scales = torch.logspace(-1, 0, len(down_block_res_samples) + 1, device=sample.device)  # 0.1 to 1.0
            scales = scales * conditioning_scale
            down_block_res_samples = [sample * scale for sample, scale in zip(down_block_res_samples, scales)]
            mid_block_res_sample = mid_block_res_sample * scales[-1]  # last one
        else:
            down_block_res_samples = [sample * conditioning_scale for sample in down_block_res_samples]
            mid_block_res_sample = mid_block_res_sample * conditioning_scale


        if not return_dict:
            return (down_block_res_samples, mid_block_res_sample)

        return ControlNetOutput(
            down_block_res_samples=down_block_res_samples, mid_block_res_sample=mid_block_res_sample
        )