init

app.py

@@ -0,0 +1,275 @@
+# Adapted from https://github.com/guoyww/AnimateDiff/blob/main/app.py
+
+# import spaces
+import os
+import json
+import torch
+import random
+
+import gradio as gr
+from glob import glob
+from omegaconf import OmegaConf
+from datetime import datetime
+from safetensors import safe_open
+
+from PIL import Image
+
+from unet2d_custom import UNet2DConditionModel
+import torch 
+from pipeline_stable_diffusion_custom import StableDiffusionPipeline 
+
+from diffusers import DDIMScheduler
+from pnp_utils import *
+import torchvision.transforms as T
+from preprocess import get_timesteps
+from preprocess import Preprocess
+        
+
+from pnp import PNP
+
+sample_idx     = 0
+
+css = """
+.toolbutton {
+    margin-buttom: 0em 0em 0em 0em;
+    max-width: 1.5em;
+    min-width: 1.5em !important;
+    height: 1.5em;
+}
+"""
+
+class AnimateController:
+    def __init__(self):
+        self.sr = 44100
+        self.save_steps = 50
+        self.device = 'cuda'
+        self.seed = 42
+        self.extract_reverse = False
+        self.save_dir = 'latents'
+        self.steps = 50 
+        self.inversion_prompt = ''
+    
+    
+        self.seed = 42
+        seed_everything(self.seed)
+        
+        self.pnp = PNP(sd_version="1.4")
+        
+        self.pnp.unet.to(self.device)
+        self.pnp.audio_projector.to(self.device)        
+        
+        
+        
+        # audio_projector_path = "ckpts/audio_projector_landscape.pth"
+        # gate_dict_path = "ckpts/landscape.pt"    
+        # self.pnp.set_audio_projector(gate_dict_path, audio_projector_path)   
+
+        
+        
+        
+    #@spaces.GPU        
+    def preprocess(self, image=None):
+
+        model_key = "CompVis/stable-diffusion-v1-4"
+        toy_scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
+        toy_scheduler.set_timesteps(self.save_steps)
+        timesteps_to_save, num_inference_steps = get_timesteps(toy_scheduler, num_inference_steps=self.save_steps,
+                                                               strength=1.0,
+                                                               device=self.device)
+        
+        save_path = os.path.join(self.save_dir + "_forward")
+        os.makedirs(save_path, exist_ok=True)
+        model = Preprocess(self.device, sd_version='1.4', hf_key=None)
+        recon_image = model.extract_latents(data_path=image,
+                                             num_steps=self.steps,
+                                             save_path=save_path,
+                                             timesteps_to_save=timesteps_to_save,
+                                             inversion_prompt=self.inversion_prompt,
+                                             extract_reverse=False)
+
+        T.ToPILImage()(recon_image[0]).save(os.path.join(save_path, f'recon.jpg'))
+
+   #@spaces.GPU    
+    def generate(self,  file=None, audio=None, prompt=None, 
+                 cfg_scale=5, image_path=None,
+                 pnp_f_t=0.8, pnp_attn_t=0.8,):
+            
+        image = self.pnp.run_pnp(
+                            n_timesteps=50, 
+                            pnp_f_t=pnp_f_t, pnp_attn_t=pnp_attn_t, 
+                            prompt=prompt,
+                            negative_prompt="",
+                            audio_path=audio, 
+                            image_path=image_path,
+                            cfg_scale=cfg_scale,
+                           )        
+        
+        return image 
+
+#     @spaces.GPU    
+#     def update_audio_model(self, audio_model_update):
+        
+#         print(f"changing ckpts audio model {audio_model_update}")
+        
+#         if audio_model_update == "Landscape Model":
+#             audio_projector_path = "ckpts/audio_projector_landscape.pth"
+#             gate_dict_path = "ckpts/landscape.pt"
+#         else:
+#             audio_projector_path = "ckpts/audio_projector_gh.pth"
+#             gate_dict_path = "ckpts/greatest_hits.pt"            
+           
+#         self.pnp.set_audio_projector(gate_dict_path, audio_projector_path)   
+#         self.pnp.changed_model = True
+        
+# #         gate_dict = torch.load(gate_dict_path)    
+        
+# #         for name, param in self.pnp.unet.named_parameters():
+# #             if "adapter" in name:
+# #                 param.data = gate_dict[name]            
+
+# #         self.pnp.audio_projector.load_state_dict(torch.load(audio_projector_path))            
+# #         self.pnp.unet.to(self.device)
+# #         self.pnp.audio_projector.to(self.device)
+        
+#         return gr.Dropdown()
+   
+controller = AnimateController()
+
+
+def ui():
+    with gr.Blocks(css=css) as demo:
+        gr.Markdown(
+            """
+            # [SonicDiffusion: Audio-Driven Image Generation and Editing with Pretrained Diffusion Models]
+            """
+        )    
+        with gr.Row():
+            audio_input = gr.Audio(sources="upload", type="filepath")
+            prompt_textbox = gr.Textbox(label="Prompt", lines=2)
+            
+        with gr.Row():
+            with gr.Column():
+                pnp_f_t = gr.Slider(label="PNP Residual Injection",    step=0.1, value=0.8, minimum=0.0,   maximum=1.0)
+                pnp_attn_t = gr.Slider(label="PNP Attention Injection", step=0.1, value=0.8, minimum=0.0,  maximum=1.0)
+                
+            with gr.Column():
+                audio_model_dropdown = gr.Dropdown(
+                    label="Select SonicDiffusion model",
+                    value="Landscape Model",
+                    choices=["Landscape Model", "Greatest Hits Model"],
+                    interactive=True,
+                )
+                
+#                 audio_model_dropdown.change(fn=controller.update_audio_model, inputs=[audio_model_dropdown], outputs=[audio_model_dropdown])
+                cfg_scale_slider = gr.Slider(label="CFG Scale", step=0.5, value=7.5, minimum=0,   maximum=20)
+                
+
+        with gr.Row():
+            preprocess_button = gr.Button(value="Preprocess", variant='primary')            
+            generate_button = gr.Button(value="Generate", variant='primary')
+
+        
+        with gr.Row():
+            with gr.Column():
+                image_input = gr.Image(label="Input Image Component", sources="upload", type="filepath")
+
+            with gr.Column():
+                output = gr.Image(label="Output Image Component", 
+                                 height=512, width=512)
+
+        with gr.Row():
+            
+            examples_img_1 = [
+                [Image.open("assets/corridor.png")],
+                [Image.open("assets/desert.png")],                
+                [Image.open("assets/forest.png")],
+                [Image.open("assets/forest_painting.png")], 
+                [Image.open("assets/golf_field.png")],
+                [Image.open("assets/human.png")],
+                [Image.open("assets/wood.png")],
+                [Image.open("assets/house.png")],
+                
+                [Image.open("assets/apple.png")],
+                [Image.open("assets/chair.png")],         
+                [Image.open("assets/hands.png")],                 
+                [Image.open("assets/pineapple.png")], 
+                [Image.open("assets/table.png")],                 
+            ]
+            gr.Examples(examples=examples_img_1,inputs=[image_input], label="Images")
+
+
+            # examples_img_2 = [
+            #     [Image.open("assets/apple.png")],
+            #     [Image.open("assets/chair.png")],         
+            #     [Image.open("assets/hands.png")],                 
+            #     [Image.open("assets/pineapple.png")], 
+            #     [Image.open("assets/table.png")],                
+            # ]
+            # gr.Examples(examples=examples,inputs=[image_input], label="Greatest Hits Images")
+            
+            examples2 = [
+                ['./assets/fire_crackling.wav'],
+                ['./assets/forest_birds.wav'],
+                ['./assets/forest_stepping_on_branches.wav'],
+                ['./assets/howling_wind.wav'],
+                ['./assets/rain.wav'],
+                ['./assets/splashing_water.wav'],
+                ['./assets/splashing_water_soft.wav'],
+                ['./assets/steps_on_snow.wav'],
+                ['./assets/thunder.wav'],
+                ['./assets/underwater.wav'],
+                ['./assets/waterfall_burble.wav'],
+                ['./assets/wind_noise_birds.wav'],
+            ]
+            
+            gr.Examples(examples=examples2,inputs=[audio_input], label="Landscape Audios")           
+
+
+            examples3 = [
+                ['./assets/cardboard.wav'],
+                ['./assets/carpet.wav'],            
+                ['./assets/ceramic.wav'],
+                ['./assets/cloth.wav'],                    
+                ['./assets/gravel.wav'],
+                ['./assets/leaf.wav'],    
+                ['./assets/metal.wav'],
+                ['./assets/plastic_bag.wav'],    
+                ['./assets/plastic.wav'],
+                ['./assets/rock.wav'],     
+                ['./assets/wood.wav'],                    
+            ]
+            gr.Examples(examples=examples3,inputs=[audio_input], label="Greatest Hits Audios")  
+        
+        
+        preprocess_button.click(
+            fn=controller.preprocess,
+            inputs=[
+                image_input
+            ],
+            outputs=output
+        )        
+    
+        
+        generate_button.click(
+            fn=controller.generate,
+            inputs=[
+                audio_model_dropdown,
+                audio_input,
+                prompt_textbox,
+                cfg_scale_slider,
+                image_input,
+                pnp_f_t,
+                pnp_attn_t,
+            ],
+            outputs=output
+        )
+            
+    
+    return demo
+
+
+
+if __name__ == "__main__":
+    demo = ui()
+    demo.launch(share=True)
+    

attention_custom.py

@@ -0,0 +1,681 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention.py
+
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.utils import deprecate, logging
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU
+from diffusers.models.attention_processor import Attention
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm
+
+
+logger = logging.get_logger(__name__)
+
+
+def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    ff_output = torch.cat(
+        [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+        dim=chunk_dim,
+    )
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+        ada_norm_bias: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+        use_adapter: bool = False,
+    ):
+        super().__init__()
+        self.only_cross_attention = only_cross_attention
+
+        # We keep these boolean flags for backward-compatibility.
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        self.use_adapter = use_adapter
+
+        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}."
+            )
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if norm_type == "ada_norm":
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif norm_type == "ada_norm_zero":
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif norm_type == "ada_norm_continuous":
+            self.norm1 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "rms_norm",
+            )
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = 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,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            if norm_type == "ada_norm":
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif norm_type == "ada_norm_continuous":
+                self.norm2 = AdaLayerNormContinuous(
+                    dim,
+                    ada_norm_continous_conditioning_embedding_dim,
+                    norm_elementwise_affine,
+                    norm_eps,
+                    ada_norm_bias,
+                    "rms_norm",
+                )
+            else:
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        if use_adapter:
+            self.attn_adapter = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+                )
+            self.norm_adapter = nn.LayerNorm(dim)
+            self.gate_adapter = nn.Parameter(torch.tensor([0.1]))
+
+
+        # 3. Feed-forward
+        if norm_type == "ada_norm_continuous":
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+
+        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm", "ada_norm_continuous"]:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+        elif norm_type == "layer_norm_i2vgen":
+            self.norm3 = None
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if norm_type == "ada_norm_single":
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        audio_context: Optional[torch.FloatTensor] = None,
+        f_multiplier: Optional[float] = 1.0,
+    ) -> torch.FloatTensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is depcrecated. `scale` will be ignored.")
+
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.norm_type == "ada_norm":
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.norm_type == "ada_norm_zero":
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.norm_type == "ada_norm_continuous":
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif self.norm_type == "ada_norm_single":
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+            norm_hidden_states = norm_hidden_states.squeeze(1)
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+        if self.norm_type == "ada_norm_zero":
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.norm_type == "ada_norm_single":
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 1.2 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.norm_type == "ada_norm":
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.norm_type == "ada_norm_single":
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.norm_type == "ada_norm_continuous":
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+
+        if self.use_adapter and audio_context is not None:
+            norm_hidden_states = self.norm_adapter(hidden_states)
+            attn_output = self.attn_adapter(norm_hidden_states, encoder_hidden_states=audio_context)*self.gate_adapter.tanh()
+            hidden_states = attn_output*f_multiplier + hidden_states
+
+
+        # 4. Feed-forward
+        # i2vgen doesn't have this norm 🤷‍♂️
+        if self.norm_type == "ada_norm_continuous":
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.norm_type == "ada_norm_single":
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.norm_type == "ada_norm_single":
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.norm_type == "ada_norm_single":
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+    ) -> torch.FloatTensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+        linear_cls = nn.Linear
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(linear_cls(inner_dim, dim_out, bias=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states

pipeline_stable_diffusion_custom.py

@@ -0,0 +1,1024 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
+
+import inspect
+from typing import Any, Callable, Dict, List, Optional, Union
+
+import torch
+from packaging import version
+from transformers import CLIPImageProcessor, CLIPTextModel, CLIPTokenizer, CLIPVisionModelWithProjection
+
+from diffusers.configuration_utils import FrozenDict
+from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
+
+from diffusers.loaders import  IPAdapterMixin, LoraLoaderMixin, TextualInversionLoaderMixin, FromSingleFileMixin
+
+from diffusers.models import AutoencoderKL, ImageProjection, UNet2DConditionModel
+from diffusers.models.lora import adjust_lora_scale_text_encoder
+from diffusers.schedulers import KarrasDiffusionSchedulers
+from diffusers.utils import (
+    USE_PEFT_BACKEND,
+    deprecate,
+    logging,
+    replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
+)
+from diffusers.utils.torch_utils import randn_tensor
+from diffusers.pipelines.pipeline_utils import DiffusionPipeline, StableDiffusionMixin
+from diffusers.pipelines.stable_diffusion.pipeline_output import StableDiffusionPipelineOutput
+from diffusers.pipelines.stable_diffusion.safety_checker import StableDiffusionSafetyChecker
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+EXAMPLE_DOC_STRING = """
+    Examples:
+        ```py
+        >>> import torch
+        >>> from diffusers import StableDiffusionPipeline
+
+        >>> pipe = StableDiffusionPipeline.from_pretrained("runwayml/stable-diffusion-v1-5", torch_dtype=torch.float16)
+        >>> pipe = pipe.to("cuda")
+
+        >>> prompt = "a photo of an astronaut riding a horse on mars"
+        >>> image = pipe(prompt).images[0]
+        ```
+"""
+
+
+def rescale_noise_cfg(noise_cfg, noise_pred_text, guidance_rescale=0.0):
+    """
+    Rescale `noise_cfg` according to `guidance_rescale`. Based on findings of [Common Diffusion Noise Schedules and
+    Sample Steps are Flawed](https://arxiv.org/pdf/2305.08891.pdf). See Section 3.4
+    """
+    std_text = noise_pred_text.std(dim=list(range(1, noise_pred_text.ndim)), keepdim=True)
+    std_cfg = noise_cfg.std(dim=list(range(1, noise_cfg.ndim)), keepdim=True)
+    # rescale the results from guidance (fixes overexposure)
+    noise_pred_rescaled = noise_cfg * (std_text / std_cfg)
+    # mix with the original results from guidance by factor guidance_rescale to avoid "plain looking" images
+    noise_cfg = guidance_rescale * noise_pred_rescaled + (1 - guidance_rescale) * noise_cfg
+    return noise_cfg
+
+
+def retrieve_timesteps(
+    scheduler,
+    num_inference_steps: Optional[int] = None,
+    device: Optional[Union[str, torch.device]] = None,
+    timesteps: Optional[List[int]] = None,
+    **kwargs,
+):
+    """
+    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
+    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.
+
+    Args:
+        scheduler (`SchedulerMixin`):
+            The scheduler to get timesteps from.
+        num_inference_steps (`int`):
+            The number of diffusion steps used when generating samples with a pre-trained model. If used,
+            `timesteps` must be `None`.
+        device (`str` or `torch.device`, *optional*):
+            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
+        timesteps (`List[int]`, *optional*):
+                Custom timesteps used to support arbitrary spacing between timesteps. If `None`, then the default
+                timestep spacing strategy of the scheduler is used. If `timesteps` is passed, `num_inference_steps`
+                must be `None`.
+
+    Returns:
+        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
+        second element is the number of inference steps.
+    """
+    if timesteps is not None:
+        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
+        if not accepts_timesteps:
+            raise ValueError(
+                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
+                f" timestep schedules. Please check whether you are using the correct scheduler."
+            )
+        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+        num_inference_steps = len(timesteps)
+    else:
+        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
+        timesteps = scheduler.timesteps
+    return timesteps, num_inference_steps
+
+
+class StableDiffusionPipeline(
+    DiffusionPipeline,
+    StableDiffusionMixin,
+    TextualInversionLoaderMixin,
+    LoraLoaderMixin,
+    IPAdapterMixin,
+    FromSingleFileMixin,
+):
+    r"""
+    Pipeline for text-to-image generation using Stable Diffusion.
+
+    This model inherits from [`DiffusionPipeline`]. Check the superclass documentation for the generic methods
+    implemented for all pipelines (downloading, saving, running on a particular device, etc.).
+
+    The pipeline also inherits the following loading methods:
+        - [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] for loading textual inversion embeddings
+        - [`~loaders.LoraLoaderMixin.load_lora_weights`] for loading LoRA weights
+        - [`~loaders.LoraLoaderMixin.save_lora_weights`] for saving LoRA weights
+        - [`~loaders.FromSingleFileMixin.from_single_file`] for loading `.ckpt` files
+        - [`~loaders.IPAdapterMixin.load_ip_adapter`] for loading IP Adapters
+
+    Args:
+        vae ([`AutoencoderKL`]):
+            Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
+        text_encoder ([`~transformers.CLIPTextModel`]):
+            Frozen text-encoder ([clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14)).
+        tokenizer ([`~transformers.CLIPTokenizer`]):
+            A `CLIPTokenizer` to tokenize text.
+        unet ([`UNet2DConditionModel`]):
+            A `UNet2DConditionModel` to denoise the encoded image latents.
+        scheduler ([`SchedulerMixin`]):
+            A scheduler to be used in combination with `unet` to denoise the encoded image latents. Can be one of
+            [`DDIMScheduler`], [`LMSDiscreteScheduler`], or [`PNDMScheduler`].
+        safety_checker ([`StableDiffusionSafetyChecker`]):
+            Classification module that estimates whether generated images could be considered offensive or harmful.
+            Please refer to the [model card](https://huggingface.co/runwayml/stable-diffusion-v1-5) for more details
+            about a model's potential harms.
+        feature_extractor ([`~transformers.CLIPImageProcessor`]):
+            A `CLIPImageProcessor` to extract features from generated images; used as inputs to the `safety_checker`.
+    """
+
+    model_cpu_offload_seq = "text_encoder->image_encoder->unet->vae"
+    _optional_components = ["safety_checker", "feature_extractor", "image_encoder"]
+    _exclude_from_cpu_offload = ["safety_checker"]
+    _callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
+
+    def __init__(
+        self,
+        vae: AutoencoderKL,
+        text_encoder: CLIPTextModel,
+        tokenizer: CLIPTokenizer,
+        unet: UNet2DConditionModel,
+        scheduler: KarrasDiffusionSchedulers,
+        safety_checker: StableDiffusionSafetyChecker,
+        feature_extractor: CLIPImageProcessor,
+        image_encoder: CLIPVisionModelWithProjection = None,
+        requires_safety_checker: bool = True,
+    ):
+        super().__init__()
+
+        if hasattr(scheduler.config, "steps_offset") and scheduler.config.steps_offset != 1:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} is outdated. `steps_offset`"
+                f" should be set to 1 instead of {scheduler.config.steps_offset}. Please make sure "
+                "to update the config accordingly as leaving `steps_offset` 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 `scheduler/scheduler_config.json`"
+                " file"
+            )
+            deprecate("steps_offset!=1", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["steps_offset"] = 1
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if hasattr(scheduler.config, "clip_sample") and scheduler.config.clip_sample is True:
+            deprecation_message = (
+                f"The configuration file of this scheduler: {scheduler} has not set the configuration `clip_sample`."
+                " `clip_sample` should be set to False in the configuration file. Please make sure to update the"
+                " config accordingly as not setting `clip_sample` in the config might lead 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 `scheduler/scheduler_config.json` file"
+            )
+            deprecate("clip_sample not set", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(scheduler.config)
+            new_config["clip_sample"] = False
+            scheduler._internal_dict = FrozenDict(new_config)
+
+        if safety_checker is None and requires_safety_checker:
+            logger.warning(
+                f"You have disabled the safety checker for {self.__class__} by passing `safety_checker=None`. Ensure"
+                " that you abide to the conditions of the Stable Diffusion license and do not expose unfiltered"
+                " results in services or applications open to the public. Both the diffusers team and Hugging Face"
+                " strongly recommend to keep the safety filter enabled in all public facing circumstances, disabling"
+                " it only for use-cases that involve analyzing network behavior or auditing its results. For more"
+                " information, please have a look at https://github.com/huggingface/diffusers/pull/254 ."
+            )
+
+        if safety_checker is not None and feature_extractor is None:
+            raise ValueError(
+                "Make sure to define a feature extractor when loading {self.__class__} if you want to use the safety"
+                " checker. If you do not want to use the safety checker, you can pass `'safety_checker=None'` instead."
+            )
+
+        is_unet_version_less_0_9_0 = hasattr(unet.config, "_diffusers_version") and version.parse(
+            version.parse(unet.config._diffusers_version).base_version
+        ) < version.parse("0.9.0.dev0")
+        is_unet_sample_size_less_64 = hasattr(unet.config, "sample_size") and unet.config.sample_size < 64
+        if is_unet_version_less_0_9_0 and is_unet_sample_size_less_64:
+            deprecation_message = (
+                "The configuration file of the unet has set the default `sample_size` to smaller than"
+                " 64 which seems highly unlikely. If your checkpoint is a fine-tuned version of any of the"
+                " following: \n- CompVis/stable-diffusion-v1-4 \n- CompVis/stable-diffusion-v1-3 \n-"
+                " CompVis/stable-diffusion-v1-2 \n- CompVis/stable-diffusion-v1-1 \n- runwayml/stable-diffusion-v1-5"
+                " \n- runwayml/stable-diffusion-inpainting \n you should change 'sample_size' to 64 in the"
+                " configuration file. Please make sure to update the config accordingly as leaving `sample_size=32`"
+                " in the config might lead 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 `unet/config.json` file"
+            )
+            deprecate("sample_size<64", "1.0.0", deprecation_message, standard_warn=False)
+            new_config = dict(unet.config)
+            new_config["sample_size"] = 64
+            unet._internal_dict = FrozenDict(new_config)
+
+        self.register_modules(
+            vae=vae,
+            text_encoder=text_encoder,
+            tokenizer=tokenizer,
+            unet=unet,
+            scheduler=scheduler,
+            safety_checker=safety_checker,
+            feature_extractor=feature_extractor,
+            image_encoder=image_encoder,
+        )
+        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.register_to_config(requires_safety_checker=requires_safety_checker)
+
+    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,
+        **kwargs,
+    ):
+        deprecation_message = "`_encode_prompt()` is deprecated and it will be removed in a future version. Use `encode_prompt()` instead. Also, be aware that the output format changed from a concatenated tensor to a tuple."
+        deprecate("_encode_prompt()", "1.0.0", deprecation_message, standard_warn=False)
+
+        prompt_embeds_tuple = 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=lora_scale,
+            **kwargs,
+        )
+
+        # concatenate for backwards comp
+        prompt_embeds = torch.cat([prompt_embeds_tuple[1], prompt_embeds_tuple[0]])
+
+        return prompt_embeds
+
+    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,
+        clip_skip: Optional[int] = 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. 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.
+            lora_scale (`float`, *optional*):
+                A LoRA scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+        """
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, LoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if not USE_PEFT_BACKEND:
+                adjust_lora_scale_text_encoder(self.text_encoder, lora_scale)
+            else:
+                scale_lora_layers(self.text_encoder, lora_scale)
+
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        if prompt_embeds is None:
+            # textual inversion: process multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                prompt = self.maybe_convert_prompt(prompt, self.tokenizer)
+
+            text_inputs = self.tokenizer(
+                prompt,
+                padding="max_length",
+                max_length=self.tokenizer.model_max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+            text_input_ids = text_inputs.input_ids
+            untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
+
+            if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(
+                text_input_ids, untruncated_ids
+            ):
+                removed_text = self.tokenizer.batch_decode(
+                    untruncated_ids[:, self.tokenizer.model_max_length - 1 : -1]
+                )
+                logger.warning(
+                    "The following part of your input was truncated because CLIP can only handle sequences up to"
+                    f" {self.tokenizer.model_max_length} tokens: {removed_text}"
+                )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = text_inputs.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            if clip_skip is None:
+                prompt_embeds = self.text_encoder(text_input_ids.to(device), attention_mask=attention_mask)
+                prompt_embeds = prompt_embeds[0]
+            else:
+                prompt_embeds = self.text_encoder(
+                    text_input_ids.to(device), attention_mask=attention_mask, output_hidden_states=True
+                )
+                # Access the `hidden_states` first, that contains a tuple of
+                # all the hidden states from the encoder layers. Then index into
+                # the tuple to access the hidden states from the desired layer.
+                prompt_embeds = prompt_embeds[-1][-(clip_skip + 1)]
+                # We also need to apply the final LayerNorm here to not mess with the
+                # representations. The `last_hidden_states` that we typically use for
+                # obtaining the final prompt representations passes through the LayerNorm
+                # layer.
+                prompt_embeds = self.text_encoder.text_model.final_layer_norm(prompt_embeds)
+
+        if self.text_encoder is not None:
+            prompt_embeds_dtype = self.text_encoder.dtype
+        elif self.unet is not None:
+            prompt_embeds_dtype = self.unet.dtype
+        else:
+            prompt_embeds_dtype = prompt_embeds.dtype
+
+        prompt_embeds = prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+        bs_embed, seq_len, _ = prompt_embeds.shape
+        # duplicate text embeddings for each generation per prompt, using mps friendly method
+        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
+        prompt_embeds = prompt_embeds.view(bs_embed * num_images_per_prompt, seq_len, -1)
+
+        # get unconditional embeddings for classifier free guidance
+        if do_classifier_free_guidance and negative_prompt_embeds is None:
+            uncond_tokens: List[str]
+            if negative_prompt is None:
+                uncond_tokens = [""] * batch_size
+            elif prompt is not None and type(prompt) is not type(negative_prompt):
+                raise TypeError(
+                    f"`negative_prompt` should be the same type to `prompt`, but got {type(negative_prompt)} !="
+                    f" {type(prompt)}."
+                )
+            elif isinstance(negative_prompt, str):
+                uncond_tokens = [negative_prompt]
+            elif batch_size != len(negative_prompt):
+                raise ValueError(
+                    f"`negative_prompt`: {negative_prompt} has batch size {len(negative_prompt)}, but `prompt`:"
+                    f" {prompt} has batch size {batch_size}. Please make sure that passed `negative_prompt` matches"
+                    " the batch size of `prompt`."
+                )
+            else:
+                uncond_tokens = negative_prompt
+
+            # textual inversion: process multi-vector tokens if necessary
+            if isinstance(self, TextualInversionLoaderMixin):
+                uncond_tokens = self.maybe_convert_prompt(uncond_tokens, self.tokenizer)
+
+            max_length = prompt_embeds.shape[1]
+            uncond_input = self.tokenizer(
+                uncond_tokens,
+                padding="max_length",
+                max_length=max_length,
+                truncation=True,
+                return_tensors="pt",
+            )
+
+            if hasattr(self.text_encoder.config, "use_attention_mask") and self.text_encoder.config.use_attention_mask:
+                attention_mask = uncond_input.attention_mask.to(device)
+            else:
+                attention_mask = None
+
+            negative_prompt_embeds = self.text_encoder(
+                uncond_input.input_ids.to(device),
+                attention_mask=attention_mask,
+            )
+            negative_prompt_embeds = negative_prompt_embeds[0]
+
+        if do_classifier_free_guidance:
+            # duplicate unconditional embeddings for each generation per prompt, using mps friendly method
+            seq_len = negative_prompt_embeds.shape[1]
+
+            negative_prompt_embeds = negative_prompt_embeds.to(dtype=prompt_embeds_dtype, device=device)
+
+            negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
+            negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
+
+        if isinstance(self, LoraLoaderMixin) and USE_PEFT_BACKEND:
+            # Retrieve the original scale by scaling back the LoRA layers
+            unscale_lora_layers(self.text_encoder, lora_scale)
+
+        return prompt_embeds, negative_prompt_embeds
+
+    def encode_image(self, image, device, num_images_per_prompt, output_hidden_states=None):
+        dtype = next(self.image_encoder.parameters()).dtype
+
+        if not isinstance(image, torch.Tensor):
+            image = self.feature_extractor(image, return_tensors="pt").pixel_values
+
+        image = image.to(device=device, dtype=dtype)
+        if output_hidden_states:
+            image_enc_hidden_states = self.image_encoder(image, output_hidden_states=True).hidden_states[-2]
+            image_enc_hidden_states = image_enc_hidden_states.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_enc_hidden_states = self.image_encoder(
+                torch.zeros_like(image), output_hidden_states=True
+            ).hidden_states[-2]
+            uncond_image_enc_hidden_states = uncond_image_enc_hidden_states.repeat_interleave(
+                num_images_per_prompt, dim=0
+            )
+            return image_enc_hidden_states, uncond_image_enc_hidden_states
+        else:
+            image_embeds = self.image_encoder(image).image_embeds
+            image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
+            uncond_image_embeds = torch.zeros_like(image_embeds)
+
+            return image_embeds, uncond_image_embeds
+
+    def prepare_ip_adapter_image_embeds(
+        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt, do_classifier_free_guidance
+    ):
+        if ip_adapter_image_embeds is None:
+            if not isinstance(ip_adapter_image, list):
+                ip_adapter_image = [ip_adapter_image]
+
+            if len(ip_adapter_image) != len(self.unet.encoder_hid_proj.image_projection_layers):
+                raise ValueError(
+                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.unet.encoder_hid_proj.image_projection_layers)} IP Adapters."
+                )
+
+            image_embeds = []
+            for single_ip_adapter_image, image_proj_layer in zip(
+                ip_adapter_image, self.unet.encoder_hid_proj.image_projection_layers
+            ):
+                output_hidden_state = not isinstance(image_proj_layer, ImageProjection)
+                single_image_embeds, single_negative_image_embeds = self.encode_image(
+                    single_ip_adapter_image, device, 1, output_hidden_state
+                )
+                single_image_embeds = torch.stack([single_image_embeds] * num_images_per_prompt, dim=0)
+                single_negative_image_embeds = torch.stack(
+                    [single_negative_image_embeds] * num_images_per_prompt, dim=0
+                )
+
+                if do_classifier_free_guidance:
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                    single_image_embeds = single_image_embeds.to(device)
+
+                image_embeds.append(single_image_embeds)
+        else:
+            repeat_dims = [1]
+            image_embeds = []
+            for single_image_embeds in ip_adapter_image_embeds:
+                if do_classifier_free_guidance:
+                    single_negative_image_embeds, single_image_embeds = single_image_embeds.chunk(2)
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                    single_negative_image_embeds = single_negative_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_negative_image_embeds.shape[1:]))
+                    )
+                    single_image_embeds = torch.cat([single_negative_image_embeds, single_image_embeds])
+                else:
+                    single_image_embeds = single_image_embeds.repeat(
+                        num_images_per_prompt, *(repeat_dims * len(single_image_embeds.shape[1:]))
+                    )
+                image_embeds.append(single_image_embeds)
+
+        return image_embeds
+
+    def run_safety_checker(self, image, device, dtype):
+        if self.safety_checker is None:
+            has_nsfw_concept = None
+        else:
+            if torch.is_tensor(image):
+                feature_extractor_input = self.image_processor.postprocess(image, output_type="pil")
+            else:
+                feature_extractor_input = self.image_processor.numpy_to_pil(image)
+            safety_checker_input = self.feature_extractor(feature_extractor_input, return_tensors="pt").to(device)
+            image, has_nsfw_concept = self.safety_checker(
+                images=image, clip_input=safety_checker_input.pixel_values.to(dtype)
+            )
+        return image, has_nsfw_concept
+
+    def decode_latents(self, latents):
+        deprecation_message = "The decode_latents method is deprecated and will be removed in 1.0.0. Please use VaeImageProcessor.postprocess(...) instead"
+        deprecate("decode_latents", "1.0.0", deprecation_message, standard_warn=False)
+
+        latents = 1 / self.vae.config.scaling_factor * latents
+        image = self.vae.decode(latents, return_dict=False)[0]
+        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
+
+    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,
+        height,
+        width,
+        callback_steps,
+        negative_prompt=None,
+        prompt_embeds=None,
+        negative_prompt_embeds=None,
+        ip_adapter_image=None,
+        ip_adapter_image_embeds=None,
+        callback_on_step_end_tensor_inputs=None,
+    ):
+        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 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 callback_on_step_end_tensor_inputs is not None and not all(
+            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
+        ):
+            raise ValueError(
+                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
+            )
+
+        if prompt is not None and prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
+                " only forward one of the two."
+            )
+        elif prompt is None and prompt_embeds is None:
+            raise ValueError(
+                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
+            )
+        elif 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 negative_prompt is not None and negative_prompt_embeds is not None:
+            raise ValueError(
+                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
+                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
+            )
+
+        if prompt_embeds is not None and negative_prompt_embeds is not None:
+            if prompt_embeds.shape != negative_prompt_embeds.shape:
+                raise ValueError(
+                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
+                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
+                    f" {negative_prompt_embeds.shape}."
+                )
+
+        if ip_adapter_image is not None and ip_adapter_image_embeds is not None:
+            raise ValueError(
+                "Provide either `ip_adapter_image` or `ip_adapter_image_embeds`. Cannot leave both `ip_adapter_image` and `ip_adapter_image_embeds` defined."
+            )
+
+        if ip_adapter_image_embeds is not None:
+            if not isinstance(ip_adapter_image_embeds, list):
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be of type `list` but is {type(ip_adapter_image_embeds)}"
+                )
+            elif ip_adapter_image_embeds[0].ndim not in [3, 4]:
+                raise ValueError(
+                    f"`ip_adapter_image_embeds` has to be a list of 3D or 4D tensors but is {ip_adapter_image_embeds[0].ndim}D"
+                )
+
+    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.latent_consistency_models.pipeline_latent_consistency_text2img.LatentConsistencyModelPipeline.get_guidance_scale_embedding
+    def get_guidance_scale_embedding(self, w, embedding_dim=512, dtype=torch.float32):
+        """
+        See https://github.com/google-research/vdm/blob/dc27b98a554f65cdc654b800da5aa1846545d41b/model_vdm.py#L298
+
+        Args:
+            timesteps (`torch.Tensor`):
+                generate embedding vectors at these timesteps
+            embedding_dim (`int`, *optional*, defaults to 512):
+                dimension of the embeddings to generate
+            dtype:
+                data type of the generated embeddings
+
+        Returns:
+            `torch.FloatTensor`: Embedding vectors with shape `(len(timesteps), embedding_dim)`
+        """
+        assert len(w.shape) == 1
+        w = w * 1000.0
+
+        half_dim = embedding_dim // 2
+        emb = torch.log(torch.tensor(10000.0)) / (half_dim - 1)
+        emb = torch.exp(torch.arange(half_dim, dtype=dtype) * -emb)
+        emb = w.to(dtype)[:, None] * emb[None, :]
+        emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+        if embedding_dim % 2 == 1:  # zero pad
+            emb = torch.nn.functional.pad(emb, (0, 1))
+        assert emb.shape == (w.shape[0], embedding_dim)
+        return emb
+
+    @property
+    def guidance_scale(self):
+        return self._guidance_scale
+
+    @property
+    def guidance_rescale(self):
+        return self._guidance_rescale
+
+    @property
+    def clip_skip(self):
+        return self._clip_skip
+
+    # 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.
+    @property
+    def do_classifier_free_guidance(self):
+        return self._guidance_scale > 1 and self.unet.config.time_cond_proj_dim is None
+
+    @property
+    def cross_attention_kwargs(self):
+        return self._cross_attention_kwargs
+
+    @property
+    def num_timesteps(self):
+        return self._num_timesteps
+
+    @property
+    def interrupt(self):
+        return self._interrupt
+
+    @torch.no_grad()
+    @replace_example_docstring(EXAMPLE_DOC_STRING)
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 50,
+        timesteps: List[int] = None,
+        guidance_scale: float = 7.5,
+        negative_prompt: Optional[Union[str, List[str]]] = None,
+        num_images_per_prompt: Optional[int] = 1,
+        eta: float = 0.0,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
+        ip_adapter_image: Optional[PipelineImageInput] = None,
+        ip_adapter_image_embeds: Optional[List[torch.FloatTensor]] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        guidance_rescale: float = 0.0,
+        clip_skip: Optional[int] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        audio_context: Optional[torch.FloatTensor] = None,
+        **kwargs,
+    ):
+        r"""
+        The call function to the pipeline for generation.
+
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide image generation. If not defined, you need to pass `prompt_embeds`.
+            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 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.5):
+                A higher guidance scale value encourages the model to generate images closely linked to the text
+                `prompt` at the expense of lower image quality. Guidance scale is enabled when `guidance_scale > 1`.
+            negative_prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide what to not include in image generation. If not defined, you need to
+                pass `negative_prompt_embeds` instead. Ignored when not using guidance (`guidance_scale < 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) from the [DDIM](https://arxiv.org/abs/2010.02502) paper. Only applies
+                to the [`~schedulers.DDIMScheduler`], and is ignored in other schedulers.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                A [`torch.Generator`](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 is 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 (prompt weighting). If not
+                provided, text embeddings are generated from the `prompt` input argument.
+            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If
+                not provided, `negative_prompt_embeds` are generated from the `negative_prompt` input argument.
+            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
+            ip_adapter_image_embeds (`List[torch.FloatTensor]`, *optional*):
+                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of IP-adapters.
+                Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. It should contain the negative image embedding
+                if `do_classifier_free_guidance` is set to `True`.
+                If not provided, embeddings are computed from the `ip_adapter_image` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generated image. Choose between `PIL.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.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the [`AttentionProcessor`] as defined in
+                [`self.processor`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            guidance_rescale (`float`, *optional*, defaults to 0.0):
+                Guidance rescale factor from [Common Diffusion Noise Schedules and Sample Steps are
+                Flawed](https://arxiv.org/pdf/2305.08891.pdf). Guidance rescale factor should fix overexposure when
+                using zero terminal SNR.
+            clip_skip (`int`, *optional*):
+                Number of layers to be skipped from CLIP while computing the prompt embeddings. A value of 1 means that
+                the output of the pre-final layer will be used for computing the prompt embeddings.
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+
+        Examples:
+
+        Returns:
+            [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] or `tuple`:
+                If `return_dict` is `True`, [`~pipelines.stable_diffusion.StableDiffusionPipelineOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is a list with the generated images and the
+                second element is a list of `bool`s indicating whether the corresponding generated image contains
+                "not-safe-for-work" (nsfw) content.
+        """
+
+        callback = kwargs.pop("callback", None)
+        callback_steps = kwargs.pop("callback_steps", None)
+
+        if callback is not None:
+            deprecate(
+                "callback",
+                "1.0.0",
+                "Passing `callback` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+            )
+        if callback_steps is not None:
+            deprecate(
+                "callback_steps",
+                "1.0.0",
+                "Passing `callback_steps` as an input argument to `__call__` is deprecated, consider using `callback_on_step_end`",
+            )
+
+        # 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
+        # to deal with lora scaling and other possible forward hooks
+
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            height,
+            width,
+            callback_steps,
+            negative_prompt,
+            prompt_embeds,
+            negative_prompt_embeds,
+            ip_adapter_image,
+            ip_adapter_image_embeds,
+            callback_on_step_end_tensor_inputs,
+        )
+
+        self._guidance_scale = guidance_scale
+        self._guidance_rescale = guidance_rescale
+        self._clip_skip = clip_skip
+        self._cross_attention_kwargs = cross_attention_kwargs
+        self._interrupt = False
+
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
+
+        device = self._execution_device
+
+        # 3. Encode input prompt
+        lora_scale = (
+            self.cross_attention_kwargs.get("scale", None) if self.cross_attention_kwargs is not None else None
+        )
+
+        prompt_embeds, negative_prompt_embeds = self.encode_prompt(
+            prompt,
+            device,
+            num_images_per_prompt,
+            self.do_classifier_free_guidance,
+            negative_prompt,
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            lora_scale=lora_scale,
+            clip_skip=self.clip_skip,
+        )
+
+        # 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
+        if self.do_classifier_free_guidance:
+            prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds])
+
+        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
+            image_embeds = self.prepare_ip_adapter_image_embeds(
+                ip_adapter_image,
+                ip_adapter_image_embeds,
+                device,
+                batch_size * num_images_per_prompt,
+                self.do_classifier_free_guidance,
+            )
+
+        # 4. Prepare timesteps
+        timesteps, num_inference_steps = retrieve_timesteps(self.scheduler, num_inference_steps, device, timesteps)
+
+        # 5. Prepare latent variables
+        num_channels_latents = self.unet.config.in_channels
+        latents = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_channels_latents,
+            height,
+            width,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
+
+        # 6. 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)
+
+        # 6.1 Add image embeds for IP-Adapter
+        added_cond_kwargs = (
+            {"image_embeds": image_embeds}
+            if (ip_adapter_image is not None or ip_adapter_image_embeds is not None)
+            else None
+        )
+
+        # 6.2 Optionally get Guidance Scale Embedding
+        timestep_cond = None
+        if self.unet.config.time_cond_proj_dim is not None:
+            guidance_scale_tensor = torch.tensor(self.guidance_scale - 1).repeat(batch_size * num_images_per_prompt)
+            timestep_cond = self.get_guidance_scale_embedding(
+                guidance_scale_tensor, embedding_dim=self.unet.config.time_cond_proj_dim
+            ).to(device=device, dtype=latents.dtype)
+
+        # 7. Denoising loop
+        num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
+        self._num_timesteps = len(timesteps)
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # expand the latents if we are doing classifier free guidance
+                latent_model_input = torch.cat([latents] * 2) if self.do_classifier_free_guidance else latents
+                latent_model_input = self.scheduler.scale_model_input(latent_model_input, t)
+
+                # predict the noise residual
+                noise_pred = self.unet(
+                    latent_model_input,
+                    t,
+                    encoder_hidden_states=prompt_embeds,
+                    timestep_cond=timestep_cond,
+                    cross_attention_kwargs=self.cross_attention_kwargs,
+                    added_cond_kwargs=added_cond_kwargs,
+                    return_dict=False,
+                    audio_context=audio_context,
+                )[0]
+
+                # perform guidance
+                if self.do_classifier_free_guidance:
+                    noise_pred_uncond, noise_pred_text = noise_pred.chunk(2)
+                    noise_pred = noise_pred_uncond + self.guidance_scale * (noise_pred_text - noise_pred_uncond)
+
+                if self.do_classifier_free_guidance and self.guidance_rescale > 0.0:
+                    # Based on 3.4. in https://arxiv.org/pdf/2305.08891.pdf
+                    noise_pred = rescale_noise_cfg(noise_pred, noise_pred_text, guidance_rescale=self.guidance_rescale)
+
+                # 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_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+                    negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+                    if callback is not None and i % callback_steps == 0:
+                        step_idx = i // getattr(self.scheduler, "order", 1)
+                        callback(step_idx, t, latents)
+
+        if not output_type == "latent":
+            image = self.vae.decode(latents / self.vae.config.scaling_factor, return_dict=False, generator=generator)[
+                0
+            ]
+            image, has_nsfw_concept = self.run_safety_checker(image, device, prompt_embeds.dtype)
+        else:
+            image = latents
+            has_nsfw_concept = None
+
+        if has_nsfw_concept is None:
+            do_denormalize = [True] * image.shape[0]
+        else:
+            do_denormalize = [not has_nsfw for has_nsfw in has_nsfw_concept]
+
+        image = self.image_processor.postprocess(image, output_type=output_type, do_denormalize=do_denormalize)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image, has_nsfw_concept)
+
+        return StableDiffusionPipelineOutput(images=image, nsfw_content_detected=has_nsfw_concept)

pnp.py

@@ -0,0 +1,279 @@
+# Adapted from https://github.com/MichalGeyer/pnp-diffusers/blob/main/pnp.py
+
+import spaces
+import glob
+import os
+from pathlib import Path
+import torch
+import torch.nn as nn
+import torchvision.transforms as T
+import argparse
+from PIL import Image
+import yaml
+from tqdm import tqdm
+from transformers import logging
+from diffusers import DDIMScheduler, StableDiffusionPipeline
+
+from pnp_utils import *
+
+from unet2d_custom import UNet2DConditionModel
+from pipeline_stable_diffusion_custom import StableDiffusionPipeline 
+from ldm.modules.encoders.audio_projector_res import Adapter
+
+# suppress partial model loading warning
+logging.set_verbosity_error()
+
+from diffusers import logging
+logging.set_verbosity_error()
+
+
+
+class PNP(nn.Module):
+    def __init__(self, sd_version="1.4", n_timesteps=50, audio_projector_ckpt_path="ckpts/audio_projector_gh.pth", 
+                adapter_ckpt_path="ckpts/greatest_hits.pt", device="cuda",
+                clap_path="CLAP/msclap",
+                clap_weights = "ckpts/CLAP_weights_2022.pth",
+                 
+                ):
+        super().__init__()
+        
+        self.device = device
+        
+        if sd_version == '2.1':
+            model_key = "stabilityai/stable-diffusion-2-1-base"
+        elif sd_version == '2.0':
+            model_key = "stabilityai/stable-diffusion-2-base"
+        elif sd_version == '1.5':
+            model_key = "runwayml/stable-diffusion-v1-5"
+        elif sd_version == '1.4':
+            model_key = "CompVis/stable-diffusion-v1-4"
+            print(f"model key is {model_key}")
+        else:
+            raise ValueError(f'Stable-diffusion version {sd_version} not supported.')
+
+        # Create SD models
+        print('Loading SD model')
+
+        
+        pipe = StableDiffusionPipeline.from_pretrained(model_key, torch_dtype=torch.float16).to("cuda")
+
+        model_id = "CompVis/stable-diffusion-v1-4"
+        self.unet = UNet2DConditionModel.from_pretrained(
+            model_id,
+            subfolder="unet",
+            use_adapter_list=[False, True, True],
+            low_cpu_mem_usage=False,
+            device_map=None
+        ).to("cuda")
+        
+
+        # gate_dict = torch.load(adapter_ckpt_path)
+        
+        # for name, param in self.unet.named_parameters():
+        #    if "adapter" in name:
+        #     param.data = gate_dict[name]
+        #unet.to(self.device);
+        
+        #pipe.unet = unet.to(self.device);        
+        
+        self.vae = pipe.vae
+        self.tokenizer = pipe.tokenizer
+        self.text_encoder = pipe.text_encoder
+#        self.unet = unet.to(self.device);
+        #pipe.unet
+                    
+        self.scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
+        self.scheduler.set_timesteps(n_timesteps, device=self.device)
+
+        self.latents_path = "latents_forward"
+        self.output_path = "PNP-results/home"
+                    
+        import os 
+        os.makedirs(self.output_path, exist_ok=True)        
+    
+        import sys
+        sys.path.append(clap_path)
+        from CLAPWrapper import CLAPWrapper
+        self.audio_encoder = CLAPWrapper(clap_weights, use_cuda=True)     
+        
+        self.audio_projector = Adapter(audio_token_count=77, transformer_layer_count=4).cuda()
+        #self.audio_projector_ckpt_path = audio_projector_ckpt_path
+        self.sr = 44100
+        # self.set_audio_projector(adapter_ckpt_path, audio_projector_ckpt_path)
+        self.text_encoder = self.text_encoder.cuda()
+
+#        self.audio_projector.load_state_dict(torch.load(audio_projector_ckpt_path))
+            
+        self.audio_projector_ckpt_path = audio_projector_ckpt_path
+        self.adapter_ckpt_path = adapter_ckpt_path
+        self.changed_model = False
+    
+    @spaces.GPU    
+    def set_audio_projector(self, adapter_ckpt_path, audio_projector_ckpt_path):    
+
+        print(f"SETTING MODEL TO {adapter_ckpt_path}")
+        gate_dict = torch.load(adapter_ckpt_path)
+        for name, param in self.unet.named_parameters():
+           if "adapter" in name:
+            param.data = gate_dict[name]        
+
+        self.unet.eval()
+        self.unet = self.unet.cuda()
+        
+        self.audio_projector.load_state_dict(torch.load(audio_projector_ckpt_path))
+        self.audio_projector.eval()    
+        self.audio_projector = self.audio_projector.cuda()
+    
+    @spaces.GPU    
+    def set_text_embeds(self, prompt, negative_prompt=""):
+        self.text_encoder = self.text_encoder.cuda()        
+        self.text_embeds = self.get_text_embeds(prompt, negative_prompt)            
+        self.pnp_guidance_embeds = self.get_text_embeds("", "").chunk(2)[0]
+    
+    @spaces.GPU
+    def set_audio_context(self, audio_path):
+
+        self.audio_projector = self.audio_projector.cuda()        
+        self.audio_encoder.clap.audio_encoder = self.audio_encoder.clap.audio_encoder.to("cuda")
+        audio_emb, _ = self.audio_encoder.get_audio_embeddings([audio_path], resample = self.sr)
+        
+        dtpye_w = self.audio_projector.audio_emb_projection[0].weight.dtype
+        device_w = self.audio_projector.audio_emb_projection[0].weight.device
+        
+        audio_emb = audio_emb.cuda()
+        audio_proj = self.audio_projector(audio_emb.unsqueeze(1))
+
+        audio_emb = torch.zeros(1, 1024).cuda()
+        audio_uc = self.audio_projector(audio_emb.unsqueeze(1))    
+
+        self.audio_context = torch.cat([audio_uc, audio_uc, audio_proj]).cuda()
+        
+        
+    @torch.no_grad()
+    @spaces.GPU    
+    def get_text_embeds(self, prompt, negative_prompt, batch_size=1):
+        # Tokenize text and get embeddings
+        text_input = self.tokenizer(prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
+                                    truncation=True, return_tensors='pt')
+        input_ids = text_input.input_ids.to("cuda")        
+        text_embeddings = self.text_encoder(input_ids)[0]
+
+        # Do the same for unconditional embeddings
+        uncond_input = self.tokenizer(negative_prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
+                                      return_tensors='pt')
+
+        uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(self.device))[0]
+
+        # Cat for final embeddings
+        text_embeddings = torch.cat([uncond_embeddings] * batch_size + [text_embeddings] * batch_size)
+        return text_embeddings
+
+    @torch.no_grad()
+    @spaces.GPU    
+    def decode_latent(self, latent):
+        self.vae = self.vae.cuda()
+        with torch.autocast(device_type='cuda', dtype=torch.float32):
+            latent = 1 / 0.18215 * latent
+            img = self.vae.decode(latent).sample
+            img = (img / 2 + 0.5).clamp(0, 1)
+        return img
+
+    #@torch.autocast(device_type='cuda', dtype=torch.float32)
+    @spaces.GPU    
+    def get_data(self, image_path):
+        self.image_path = image_path
+        # load image
+        image = Image.open(image_path).convert('RGB') 
+        image = image.resize((512, 512), resample=Image.Resampling.LANCZOS)
+        image = T.ToTensor()(image).to(self.device)
+        
+        # get noise
+        latents_path = os.path.join(self.latents_path, f'noisy_latents_{self.scheduler.timesteps[0]}.pt')
+        noisy_latent = torch.load(latents_path).to(self.device)
+        return image, noisy_latent
+
+    @torch.no_grad()
+    @spaces.GPU    
+    def denoise_step(self, x, t, guidance_scale):
+        # register the time step and features in pnp injection modules
+        source_latents = load_source_latents_t(t, os.path.join(self.latents_path))
+        latent_model_input = torch.cat([source_latents] + ([x] * 2))
+
+        register_time(self, t.item())
+
+        # compute text embeddings
+        text_embed_input = torch.cat([self.pnp_guidance_embeds, self.text_embeds], dim=0)
+
+        # apply the denoising network
+        noise_pred = self.unet(latent_model_input, t, 
+                               encoder_hidden_states=text_embed_input,
+                              audio_context=self.audio_context)['sample']
+        
+        # perform guidance
+        _, noise_pred_uncond, noise_pred_cond = noise_pred.chunk(3)
+        noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
+
+        # compute the denoising step with the reference model
+        denoised_latent = self.scheduler.step(noise_pred, t, x)['prev_sample']
+        return denoised_latent
+
+    @spaces.GPU        
+    def init_pnp(self, conv_injection_t, qk_injection_t):
+        self.qk_injection_timesteps = self.scheduler.timesteps[:qk_injection_t] if qk_injection_t >= 0 else []
+        self.conv_injection_timesteps = self.scheduler.timesteps[:conv_injection_t] if conv_injection_t >= 0 else []
+        register_attention_control_efficient(self, self.qk_injection_timesteps)
+        register_conv_control_efficient(self, self.conv_injection_timesteps)
+
+    @spaces.GPU            
+    def run_pnp(self, n_timesteps=50, pnp_f_t=0.5, pnp_attn_t=0.5,
+                prompt="", negative_prompt="", 
+                audio_path="", image_path="", 
+                cfg_scale=5):
+            
+#         if not self.changed_model:
+#             self.set_audio_projector(self.adapter_ckpt_path, self.audio_projector_ckpt_path)    
+        
+        self.audio_projector = self.audio_projector.cuda()
+        
+        self.set_text_embeds(prompt)     
+        self.set_audio_context(audio_path=audio_path)
+        self.image, self.eps = self.get_data(image_path=image_path)  
+        
+        self.unet = self.unet.cuda()
+                    
+        pnp_f_t = int(n_timesteps * pnp_f_t)
+        pnp_attn_t = int(n_timesteps * pnp_attn_t)
+        self.init_pnp(conv_injection_t=pnp_f_t, qk_injection_t=pnp_attn_t)
+        
+        edited_img = self.sample_loop(self.eps, cfg_scale=cfg_scale)
+
+        return T.ToPILImage()(edited_img[0])      
+        
+    @spaces.GPU   
+    def sample_loop(self, x, cfg_scale):
+        with torch.autocast(device_type='cuda', dtype=torch.float32):
+            for i, t in enumerate(tqdm(self.scheduler.timesteps, desc="Sampling")):
+                x = self.denoise_step(x, t, cfg_scale)
+
+            decoded_latent = self.decode_latent(x)
+            T.ToPILImage()(decoded_latent[0]).save(f'{self.output_path}/output.png')
+                
+        return decoded_latent
+
+
+if __name__ == '__main__':
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--config_path', type=str, default='config_pnp.yaml')
+    
+    opt = parser.parse_args()
+    with open(opt.config_path, "r") as f:
+        config = yaml.safe_load(f)
+    os.makedirs(config["output_path"], exist_ok=True)
+    
+    with open(os.path.join(config["output_path"], "config.yaml"), "w") as f:
+        yaml.dump(config, f)
+    
+    seed_everything(config["seed"])
+    print(config)
+    pnp = PNP(config)
+    temp = pnp.run_pnp()

pnp_utils.py

@@ -0,0 +1,169 @@
+# Adapted from https://github.com/MichalGeyer/pnp-diffusers/blob/main/pnp_utils.py
+
+import torch
+import os
+import random
+import numpy as np
+
+def seed_everything(seed):
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    random.seed(seed)
+    np.random.seed(seed)
+
+def register_time(model, t):
+    
+    
+    conv_res_dict = {1: [0, 1, 2], 2: [0, 1, 2], 3: [0, 1, 2]}
+    for res in conv_res_dict:
+        for block in conv_res_dict[res]:    
+            conv_module = model.unet.up_blocks[res].resnets[block]
+            setattr(conv_module, 't', t)
+        
+    down_res_dict = {0: [0, 1], 1: [0, 1], 2: [0, 1]}
+    up_res_dict = {1: [0, 1, 2], 2: [0, 1, 2], 3: [0, 1, 2]}
+    for res in up_res_dict:
+        for block in up_res_dict[res]:
+            module = model.unet.up_blocks[res].attentions[block].transformer_blocks[0].attn1
+            setattr(module, 't', t)
+    for res in down_res_dict:
+        for block in down_res_dict[res]:
+            module = model.unet.down_blocks[res].attentions[block].transformer_blocks[0].attn1
+            setattr(module, 't', t)
+    module = model.unet.mid_block.attentions[0].transformer_blocks[0].attn1
+    setattr(module, 't', t)
+
+
+def load_source_latents_t(t, latents_path):
+    latents_t_path = os.path.join(latents_path, f'noisy_latents_{t}.pt')
+    assert os.path.exists(latents_t_path), f'Missing latents at t {t} path {latents_t_path}'
+    latents = torch.load(latents_t_path)
+    return latents
+ 
+def register_attention_control_efficient(model, injection_schedule):
+    def sa_forward(self):
+        to_out = self.to_out
+        if type(to_out) is torch.nn.modules.container.ModuleList:
+            to_out = self.to_out[0]
+        else:
+            to_out = self.to_out
+
+        def forward(x, encoder_hidden_states=None, attention_mask=None):
+            batch_size, sequence_length, dim = x.shape
+            h = self.heads
+
+            is_cross = encoder_hidden_states is not None
+            encoder_hidden_states = encoder_hidden_states if is_cross else x
+            if not is_cross and self.injection_schedule is not None and (
+                    self.t in self.injection_schedule or self.t == 1000):
+                q = self.to_q(x)
+                k = self.to_k(encoder_hidden_states)
+
+                source_batch_size = int(q.shape[0] // 3)
+                # inject unconditional
+                q[source_batch_size:2 * source_batch_size] = q[:source_batch_size]
+                k[source_batch_size:2 * source_batch_size] = k[:source_batch_size]
+                # inject conditional
+                q[2 * source_batch_size:] = q[:source_batch_size]
+                k[2 * source_batch_size:] = k[:source_batch_size]
+
+                q = self.head_to_batch_dim(q)
+                k = self.head_to_batch_dim(k)
+            else:
+                q = self.to_q(x)
+                k = self.to_k(encoder_hidden_states)
+                q = self.head_to_batch_dim(q)
+                k = self.head_to_batch_dim(k)
+
+            v = self.to_v(encoder_hidden_states)
+            v = self.head_to_batch_dim(v)
+
+            sim = torch.einsum("b i d, b j d -> b i j", q, k) * self.scale
+
+            if attention_mask is not None:
+                attention_mask = attention_mask.reshape(batch_size, -1)
+                max_neg_value = -torch.finfo(sim.dtype).max
+                attention_mask = attention_mask[:, None, :].repeat(h, 1, 1)
+                sim.masked_fill_(~attention_mask, max_neg_value)
+
+            # attention, what we cannot get enough of
+            attn = sim.softmax(dim=-1)
+            out = torch.einsum("b i j, b j d -> b i d", attn, v)
+            out = self.batch_to_head_dim(out)
+
+            return to_out(out)
+
+        return forward
+
+    
+    res_dict = {1: [1, 2], 2: [0, 1, 2], 3: [0, 1, 2]}  # we are injecting attention in blocks 4 - 11 of the decoder, so not in the first block of the lowest resolution
+    for res in res_dict:
+        for block in res_dict[res]:
+            module = model.unet.up_blocks[res].attentions[block].transformer_blocks[0].attn1
+            module.forward = sa_forward(module)
+            setattr(module, 'injection_schedule', injection_schedule)
+
+
+def register_conv_control_efficient(model, injection_schedule):
+    def conv_forward(self):
+        def forward(input_tensor, temb):
+            hidden_states = input_tensor
+
+            hidden_states = self.norm1(hidden_states)
+            hidden_states = self.nonlinearity(hidden_states)
+
+            if self.upsample is not None:
+                # upsample_nearest_nhwc fails with large batch sizes. see https://github.com/huggingface/diffusers/issues/984
+                if hidden_states.shape[0] >= 64:
+                    input_tensor = input_tensor.contiguous()
+                    hidden_states = hidden_states.contiguous()
+                input_tensor = self.upsample(input_tensor)
+                hidden_states = self.upsample(hidden_states)
+            elif self.downsample is not None:
+                input_tensor = self.downsample(input_tensor)
+                hidden_states = self.downsample(hidden_states)
+
+            hidden_states = self.conv1(hidden_states)
+
+            if temb is not None:
+                temb = self.time_emb_proj(self.nonlinearity(temb))[:, :, None, None]
+
+            if temb is not None and self.time_embedding_norm == "default":
+                hidden_states = hidden_states + temb
+
+            hidden_states = self.norm2(hidden_states)
+
+            if temb is not None and self.time_embedding_norm == "scale_shift":
+                scale, shift = torch.chunk(temb, 2, dim=1)
+                hidden_states = hidden_states * (1 + scale) + shift
+
+            hidden_states = self.nonlinearity(hidden_states)
+
+            hidden_states = self.dropout(hidden_states)
+            hidden_states = self.conv2(hidden_states)
+            if self.injection_schedule is not None and (self.t in self.injection_schedule or self.t == 1000):
+                source_batch_size = int(hidden_states.shape[0] // 3)
+                # inject unconditional
+                hidden_states[source_batch_size:2 * source_batch_size] = hidden_states[:source_batch_size]
+                # inject conditional
+                hidden_states[2 * source_batch_size:] = hidden_states[:source_batch_size]
+
+            if self.conv_shortcut is not None:
+                input_tensor = self.conv_shortcut(input_tensor)
+
+            output_tensor = (input_tensor + hidden_states) / self.output_scale_factor
+
+            return output_tensor
+
+        return forward
+
+    conv_res_dict = {1: [1, 2]}
+    
+    for res in conv_res_dict:
+        for block in conv_res_dict[res]:    
+            conv_module = model.unet.up_blocks[res].resnets[block]
+            conv_module.forward = conv_forward(conv_module)
+            setattr(conv_module, 'injection_schedule', injection_schedule)    
+    
+
+    

preprocess.py

@@ -0,0 +1,228 @@
+# Adapted from https://github.com/MichalGeyer/pnp-diffusers/blob/main/preprocess.py
+
+from transformers import CLIPTextModel, CLIPTokenizer, logging
+from diffusers import AutoencoderKL, UNet2DConditionModel, DDIMScheduler
+
+# suppress partial model loading warning
+logging.set_verbosity_error()
+
+import os
+from PIL import Image
+from tqdm import tqdm, trange
+import torch
+import torch.nn as nn
+import argparse
+from pathlib import Path
+from pnp_utils import *
+import torchvision.transforms as T
+
+
+def get_timesteps(scheduler, num_inference_steps, strength, device):
+    # get the original timestep using init_timestep
+    init_timestep = min(int(num_inference_steps * strength), num_inference_steps)
+
+    t_start = max(num_inference_steps - init_timestep, 0)
+    timesteps = scheduler.timesteps[t_start:]
+
+    return timesteps, num_inference_steps - t_start
+
+
+class Preprocess(nn.Module):
+    def __init__(self, device, sd_version='2.0', hf_key=None):
+        super().__init__()
+
+        self.device = device
+        self.sd_version = sd_version
+        self.use_depth = False
+
+        print(f'[INFO] loading stable diffusion...')
+        if hf_key is not None:
+            print(f'[INFO] using hugging face custom model key: {hf_key}')
+            model_key = hf_key
+        elif self.sd_version == '2.1':
+            model_key = "stabilityai/stable-diffusion-2-1-base"
+        elif self.sd_version == '2.0':
+            model_key = "stabilityai/stable-diffusion-2-base"
+        elif self.sd_version == '1.5':
+            model_key = "runwayml/stable-diffusion-v1-5"
+        elif self.sd_version == 'depth':
+            model_key = "stabilityai/stable-diffusion-2-depth"
+            self.use_depth = True
+        elif self.sd_version == '1.4':
+            model_key = "CompVis/stable-diffusion-v1-4"            
+            
+        else:
+            raise ValueError(f'Stable-diffusion version {self.sd_version} not supported.')
+
+        # Create model
+        self.vae = AutoencoderKL.from_pretrained(model_key, subfolder="vae", revision="fp16",
+                                                 torch_dtype=torch.float16).to(self.device)
+        self.tokenizer = CLIPTokenizer.from_pretrained(model_key, subfolder="tokenizer")
+        self.text_encoder = CLIPTextModel.from_pretrained(model_key, subfolder="text_encoder", revision="fp16",
+                                                          torch_dtype=torch.float16).to(self.device)
+        self.unet = UNet2DConditionModel.from_pretrained(model_key, subfolder="unet", revision="fp16",
+                                                         torch_dtype=torch.float16).to(self.device)
+        self.scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
+        print(f'[INFO] loaded stable diffusion!')
+
+        self.inversion_func = self.ddim_inversion
+
+    @torch.no_grad()
+    def get_text_embeds(self, prompt, negative_prompt, device="cuda"):
+        text_input = self.tokenizer(prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
+                                    truncation=True, return_tensors='pt')
+        text_embeddings = self.text_encoder(text_input.input_ids.to(device))[0]
+        uncond_input = self.tokenizer(negative_prompt, padding='max_length', max_length=self.tokenizer.model_max_length,
+                                      return_tensors='pt')
+        uncond_embeddings = self.text_encoder(uncond_input.input_ids.to(device))[0]
+        text_embeddings = torch.cat([uncond_embeddings, text_embeddings])
+        return text_embeddings
+
+    @torch.no_grad()
+    def decode_latents(self, latents):
+        with torch.autocast(device_type='cuda', dtype=torch.float32):
+            latents = 1 / 0.18215 * latents
+            imgs = self.vae.decode(latents).sample
+            imgs = (imgs / 2 + 0.5).clamp(0, 1)
+        return imgs
+
+    def load_img(self, image_path):
+        image_pil = T.Resize(512)(Image.open(image_path).convert("RGB"))
+        image = T.ToTensor()(image_pil).unsqueeze(0).to(self.device)
+        return image
+
+    @torch.no_grad()
+    def encode_imgs(self, imgs):
+        with torch.autocast(device_type='cuda', dtype=torch.float32):
+            imgs = 2 * imgs - 1
+            posterior = self.vae.encode(imgs).latent_dist
+            latents = posterior.mean * 0.18215
+        return latents
+
+    @torch.no_grad()
+    def ddim_inversion(self, cond, latent, save_path, save_latents=True,
+                                timesteps_to_save=None):
+        timesteps = reversed(self.scheduler.timesteps)
+        with torch.autocast(device_type='cuda', dtype=torch.float32):
+            for i, t in enumerate(tqdm(timesteps)):
+                cond_batch = cond.repeat(latent.shape[0], 1, 1)
+
+                alpha_prod_t = self.scheduler.alphas_cumprod[t]
+                alpha_prod_t_prev = (
+                    self.scheduler.alphas_cumprod[timesteps[i - 1]]
+                    if i > 0 else self.scheduler.final_alpha_cumprod
+                )
+
+                mu = alpha_prod_t ** 0.5
+                mu_prev = alpha_prod_t_prev ** 0.5
+                sigma = (1 - alpha_prod_t) ** 0.5
+                sigma_prev = (1 - alpha_prod_t_prev) ** 0.5
+
+                eps = self.unet(latent, t, encoder_hidden_states=cond_batch).sample
+
+                pred_x0 = (latent - sigma_prev * eps) / mu_prev
+                latent = mu * pred_x0 + sigma * eps
+                if save_latents:
+                    torch.save(latent, os.path.join(save_path, f'noisy_latents_{t}.pt'))
+        torch.save(latent, os.path.join(save_path, f'noisy_latents_{t}.pt'))
+        return latent
+
+    @torch.no_grad()
+    def ddim_sample(self, x, cond, save_path, save_latents=False, timesteps_to_save=None):
+        timesteps = self.scheduler.timesteps
+        with torch.autocast(device_type='cuda', dtype=torch.float32):
+            for i, t in enumerate(tqdm(timesteps)):
+                    cond_batch = cond.repeat(x.shape[0], 1, 1)
+                    alpha_prod_t = self.scheduler.alphas_cumprod[t]
+                    alpha_prod_t_prev = (
+                        self.scheduler.alphas_cumprod[timesteps[i + 1]]
+                        if i < len(timesteps) - 1
+                        else self.scheduler.final_alpha_cumprod
+                    )
+                    mu = alpha_prod_t ** 0.5
+                    sigma = (1 - alpha_prod_t) ** 0.5
+                    mu_prev = alpha_prod_t_prev ** 0.5
+                    sigma_prev = (1 - alpha_prod_t_prev) ** 0.5
+
+                    eps = self.unet(x, t, encoder_hidden_states=cond_batch).sample
+
+                    pred_x0 = (x - sigma * eps) / mu
+                    x = mu_prev * pred_x0 + sigma_prev * eps
+
+            if save_latents:
+                torch.save(x, os.path.join(save_path, f'noisy_latents_{t}.pt'))
+        return x
+
+    @torch.no_grad()
+    def extract_latents(self, num_steps, data_path, save_path, timesteps_to_save,
+                        inversion_prompt='', extract_reverse=False):
+        self.scheduler.set_timesteps(num_steps)
+
+        cond = self.get_text_embeds(inversion_prompt, "")[1].unsqueeze(0)
+        image = self.load_img(data_path)
+        latent = self.encode_imgs(image)
+
+        inverted_x = self.inversion_func(cond, latent, save_path, save_latents=not extract_reverse,
+                                         timesteps_to_save=timesteps_to_save)
+        latent_reconstruction = self.ddim_sample(inverted_x, cond, save_path, save_latents=extract_reverse,
+                                                 timesteps_to_save=timesteps_to_save)
+        rgb_reconstruction = self.decode_latents(latent_reconstruction)
+
+        return rgb_reconstruction  # , latent_reconstruction
+
+
+def run(opt):
+    device = 'cuda'
+    # timesteps to save
+    if opt.sd_version == '2.1':
+        model_key = "stabilityai/stable-diffusion-2-1-base"
+    elif opt.sd_version == '2.0':
+        model_key = "stabilityai/stable-diffusion-2-base"
+    elif opt.sd_version == '1.5':
+        model_key = "runwayml/stable-diffusion-v1-5"
+    elif opt.sd_version == 'depth':
+        model_key = "stabilityai/stable-diffusion-2-depth"
+    elif opt.sd_version == '1.4':
+        model_key = "CompVis/stable-diffusion-v1-4"
+        
+    toy_scheduler = DDIMScheduler.from_pretrained(model_key, subfolder="scheduler")
+    toy_scheduler.set_timesteps(opt.save_steps)
+    timesteps_to_save, num_inference_steps = get_timesteps(toy_scheduler, num_inference_steps=opt.save_steps,
+                                                           strength=1.0,
+                                                           device=device)
+
+    seed_everything(opt.seed)
+
+    extraction_path_prefix = "_reverse" if opt.extract_reverse else "_forward"
+    save_path = os.path.join(opt.save_dir + extraction_path_prefix, os.path.splitext(os.path.basename(opt.data_path))[0])
+    os.makedirs(save_path, exist_ok=True)
+
+    model = Preprocess(device, sd_version=opt.sd_version, hf_key=None)
+    recon_image = model.extract_latents(data_path=opt.data_path,
+                                         num_steps=opt.steps,
+                                         save_path=save_path,
+                                         timesteps_to_save=timesteps_to_save,
+                                         inversion_prompt=opt.inversion_prompt,
+                                         extract_reverse=opt.extract_reverse)
+
+    T.ToPILImage()(recon_image[0]).save(os.path.join(save_path, f'recon.jpg'))
+
+
+if __name__ == "__main__":
+    device = 'cuda'
+    parser = argparse.ArgumentParser()
+    parser.add_argument('--data_path', type=str,
+                        default='data/source_2.png')
+    parser.add_argument('--save_dir', type=str, default='latents')
+    parser.add_argument('--sd_version', type=str, default='1.4', choices=['1.5', '2.0', '2.1', '1.4'],
+                        help="stable diffusion version")
+    
+    parser.add_argument('--seed', type=int, default=1)
+    parser.add_argument('--steps', type=int, default=50)
+    parser.add_argument('--save-steps', type=int, default=1000)
+    parser.add_argument('--inversion_prompt', type=str, default='')
+    parser.add_argument('--extract-reverse', default=False, action='store_true', help="extract features during the denoising process")
+    
+    opt = parser.parse_args()
+    run(opt)
+    

transformer_2d_custom.py

@@ -0,0 +1,453 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/transformer_2d.py
+
+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.utils import BaseOutput, deprecate, is_torch_version, logging
+from attention_custom import BasicTransformerBlock
+
+from diffusers.models.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
+from diffusers.models.modeling_utils import ModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+@dataclass
+class Transformer2DModelOutput(BaseOutput):
+    """
+    The output of [`Transformer2DModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)` or `(batch size, num_vector_embeds - 1, num_latent_pixels)` if [`Transformer2DModel`] is discrete):
+            The hidden states output conditioned on the `encoder_hidden_states` input. If discrete, returns probability
+            distributions for the unnoised latent pixels.
+    """
+
+    sample: torch.FloatTensor
+
+
+class Transformer2DModel(ModelMixin, ConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+        interpolation_scale: float = None,
+        use_adapter: bool = False,
+    ):
+        super().__init__()
+        if patch_size is not None:
+            if norm_type not in ["ada_norm", "ada_norm_zero", "ada_norm_single"]:
+                raise NotImplementedError(
+                    f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'."
+                )
+            elif norm_type in ["ada_norm", "ada_norm_zero"] and num_embeds_ada_norm is None:
+                raise ValueError(
+                    f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None."
+                )
+
+        self.use_linear_projection = use_linear_projection
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        inner_dim = num_attention_heads * attention_head_dim
+
+        conv_cls = nn.Conv2d
+        linear_cls = nn.Linear
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`.Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        # 2. Define input layers
+        if self.is_input_continuous:
+            self.in_channels = in_channels
+
+            self.norm = torch.nn.GroupNorm(num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True)
+            if use_linear_projection:
+                self.proj_in = linear_cls(in_channels, inner_dim)
+            else:
+                self.proj_in = conv_cls(in_channels, inner_dim, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            assert sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+            assert num_vector_embeds is not None, "Transformer2DModel over discrete input must provide num_embed"
+
+            self.height = sample_size
+            self.width = sample_size
+            self.num_vector_embeds = num_vector_embeds
+            self.num_latent_pixels = self.height * self.width
+
+            self.latent_image_embedding = ImagePositionalEmbeddings(
+                num_embed=num_vector_embeds, embed_dim=inner_dim, height=self.height, width=self.width
+            )
+        elif self.is_input_patches:
+            assert sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+            self.height = sample_size
+            self.width = sample_size
+
+            self.patch_size = patch_size
+            interpolation_scale = (
+                interpolation_scale if interpolation_scale is not None else max(self.config.sample_size // 64, 1)
+            )
+            self.pos_embed = PatchEmbed(
+                height=sample_size,
+                width=sample_size,
+                patch_size=patch_size,
+                in_channels=in_channels,
+                embed_dim=inner_dim,
+                interpolation_scale=interpolation_scale,
+            )
+
+        # 3. Define transformers blocks
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    inner_dim,
+                    num_attention_heads,
+                    attention_head_dim,
+                    dropout=dropout,
+                    cross_attention_dim=cross_attention_dim,
+                    activation_fn=activation_fn,
+                    num_embeds_ada_norm=num_embeds_ada_norm,
+                    attention_bias=attention_bias,
+                    only_cross_attention=only_cross_attention,
+                    double_self_attention=double_self_attention,
+                    upcast_attention=upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=norm_elementwise_affine,
+                    norm_eps=norm_eps,
+                    attention_type=attention_type,
+                    use_adapter=use_adapter,
+                )
+                for d in range(num_layers)
+            ]
+        )
+
+        # 4. Define output layers
+        self.out_channels = in_channels if out_channels is None else out_channels
+        if self.is_input_continuous:
+            # TODO: should use out_channels for continuous projections
+            if use_linear_projection:
+                self.proj_out = linear_cls(inner_dim, in_channels)
+            else:
+                self.proj_out = conv_cls(inner_dim, in_channels, kernel_size=1, stride=1, padding=0)
+        elif self.is_input_vectorized:
+            self.norm_out = nn.LayerNorm(inner_dim)
+            self.out = nn.Linear(inner_dim, self.num_vector_embeds - 1)
+        elif self.is_input_patches and norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(inner_dim, 2 * inner_dim)
+            self.proj_out_2 = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+        elif self.is_input_patches and norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, inner_dim) / inner_dim**0.5)
+            self.proj_out = nn.Linear(inner_dim, patch_size * patch_size * self.out_channels)
+
+        # 5. PixArt-Alpha blocks.
+        self.adaln_single = None
+        self.use_additional_conditions = False
+        if norm_type == "ada_norm_single":
+            self.use_additional_conditions = self.config.sample_size == 128
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(inner_dim, use_additional_conditions=self.use_additional_conditions)
+
+        self.caption_projection = None
+        if caption_channels is not None:
+            self.caption_projection = PixArtAlphaTextProjection(in_features=caption_channels, hidden_size=inner_dim)
+
+        self.gradient_checkpointing = False
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        audio_context: Optional[torch.Tensor] = None,
+        f_multiplier: Optional[float] = 1.0
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.FloatTensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.FloatTensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.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.
+        """
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is depcrecated. `scale` will be ignored.")
+        # 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:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            hidden_states = self.pos_embed(hidden_states)
+
+            if self.adaln_single is not None:
+                if self.use_additional_conditions and added_cond_kwargs is None:
+                    raise ValueError(
+                        "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                    )
+                batch_size = hidden_states.shape[0]
+                timestep, embedded_timestep = self.adaln_single(
+                    timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+                )
+
+        # 2. Blocks
+        if self.caption_projection is not None:
+            batch_size = hidden_states.shape[0]
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        for block in self.transformer_blocks:
+            if self.training and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    audio_context,
+                    f_multiplier,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                    audio_context=audio_context,
+                    f_multiplier=f_multiplier,                    
+                )
+
+        # 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()
+
+        if self.is_input_patches:
+            if self.config.norm_type != "ada_norm_single":
+                conditioning = self.transformer_blocks[0].norm1.emb(
+                    timestep, class_labels, hidden_dtype=hidden_states.dtype
+                )
+                shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+                hidden_states = self.proj_out_2(hidden_states)
+            elif self.config.norm_type == "ada_norm_single":
+                shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+                hidden_states = self.norm_out(hidden_states)
+                # Modulation
+                hidden_states = hidden_states * (1 + scale) + shift
+                hidden_states = self.proj_out(hidden_states)
+                hidden_states = hidden_states.squeeze(1)
+
+            # unpatchify
+            if self.adaln_single is None:
+                height = width = int(hidden_states.shape[1] ** 0.5)
+            hidden_states = hidden_states.reshape(
+                shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+            )
+            hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+            output = hidden_states.reshape(
+                shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+            )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)

unet2d_custom.py

@@ -0,0 +1,1314 @@
+# Adapted from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/unet_2d_blocks.py
+
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+#from diffusers.loaders import UNet2DConditionLoadersMixin
+
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.activations import get_activation
+
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    #GLIGENTextBoundingboxProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+
+from unet_2d_blocks_custom import (
+    get_down_block,
+    get_mid_block,
+    get_up_block,
+)
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.FloatTensor = None
+
+
+class UNet2DConditionModel(ModelMixin, ConfigMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[int] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: float = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads: int = 64,
+        use_adapter_list: list = [False, False, False],
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        self._check_config(
+            down_block_types=down_block_types,
+            up_block_types=up_block_types,
+            only_cross_attention=only_cross_attention,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            cross_attention_dim=cross_attention_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+        )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        time_embed_dim, timestep_input_dim = self._set_time_proj(
+            time_embedding_type,
+            block_out_channels=block_out_channels,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            time_embedding_dim=time_embedding_dim,
+        )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        self._set_encoder_hid_proj(
+            encoder_hid_dim_type,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+        )
+
+        # class embedding
+        self._set_class_embedding(
+            class_embed_type,
+            act_fn=act_fn,
+            num_class_embeds=num_class_embeds,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+            timestep_input_dim=timestep_input_dim,
+        )
+
+        self._set_add_embedding(
+            addition_embed_type,
+            addition_embed_type_num_heads=addition_embed_type_num_heads,
+            addition_time_embed_dim=addition_time_embed_dim,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+        )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+                use_adapter=use_adapter_list[0],
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = get_mid_block(
+            mid_block_type,
+            temb_channels=blocks_time_embed_dim,
+            in_channels=block_out_channels[-1],
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            output_scale_factor=mid_block_scale_factor,
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            num_attention_heads=num_attention_heads[-1],
+            cross_attention_dim=cross_attention_dim[-1],
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            mid_block_only_cross_attention=mid_block_only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+            resnet_skip_time_act=resnet_skip_time_act,
+            cross_attention_norm=cross_attention_norm,
+            attention_head_dim=attention_head_dim[-1],
+            dropout=dropout,
+            use_adapter=use_adapter_list[1],
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+                use_adapter=use_adapter_list[2],
+            )
+            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
+        )
+
+        self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
+
+    def _check_config(
+        self,
+        down_block_types: Tuple[str],
+        up_block_types: Tuple[str],
+        only_cross_attention: Union[bool, Tuple[bool]],
+        block_out_channels: Tuple[int],
+        layers_per_block: Union[int, Tuple[int]],
+        cross_attention_dim: Union[int, Tuple[int]],
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
+        reverse_transformer_layers_per_block: bool,
+        attention_head_dim: int,
+        num_attention_heads: Optional[Union[int, Tuple[int]]],
+    ):
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+    def _set_time_proj(
+        self,
+        time_embedding_type: str,
+        block_out_channels: int,
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        time_embedding_dim: int,
+    ) -> Tuple[int, int]:
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        return time_embed_dim, timestep_input_dim
+
+    def _set_encoder_hid_proj(
+        self,
+        encoder_hid_dim_type: Optional[str],
+        cross_attention_dim: Union[int, Tuple[int]],
+        encoder_hid_dim: Optional[int],
+    ):
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.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
+
+    def _set_class_embedding(
+        self,
+        class_embed_type: Optional[str],
+        act_fn: str,
+        num_class_embeds: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+        timestep_input_dim: int,
+    ):
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(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
+
+    def _set_add_embedding(
+        self,
+        addition_embed_type: str,
+        addition_embed_type_num_heads: int,
+        addition_time_embed_dim: Optional[int],
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        cross_attention_dim: Optional[int],
+        encoder_hid_dim: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+    ):
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (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'.")
+
+    def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, tuple) or isinstance(cross_attention_dim, list):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = GLIGENTextBoundingboxProjection(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: 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 ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query,
+        key, value) are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def unload_lora(self):
+        """Unloads LoRA weights."""
+        deprecate(
+            "unload_lora",
+            "0.28.0",
+            "Calling `unload_lora()` is deprecated and will be removed in a future version. Please install `peft` and then call `disable_adapters().",
+        )
+        for module in self.modules():
+            if hasattr(module, "set_lora_layer"):
+                module.set_lora_layer(None)
+
+    def get_time_embed(
+        self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
+    ) -> Optional[torch.Tensor]:
+        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)
+        return t_emb
+
+    def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
+        class_emb = None
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.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)
+        return class_emb
+
+    def get_aug_embed(
+        self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> Optional[torch.Tensor]:
+        aug_emb = None
+        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 = self.add_embedding(image_embs, hint)
+        return aug_emb
+
+    def process_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> torch.Tensor:
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kadinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in  `added_conditions`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds)
+            encoder_hidden_states = (encoder_hidden_states, image_embeds)
+        return encoder_hidden_states
+
+    def forward(
+        self,
+        sample: torch.FloatTensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+        audio_context: Optional[torch.Tensor] = None,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.FloatTensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.FloatTensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.FloatTensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+            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.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned, otherwise
+                a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+        emb = self.time_embedding(t_emb, timestep_cond)
+        aug_emb = None
+
+        class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+        if class_emb is not None:
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        aug_emb = self.get_aug_embed(
+            emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+        if self.config.addition_embed_type == "image_hint":
+            aug_emb, hint = aug_emb
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
+        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
+        if cross_attention_kwargs is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            lora_scale = cross_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    audio_context=audio_context,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    audio_context=audio_context,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                    audio_context=audio_context,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)