stylestudio

app.py

@@ -0,0 +1,238 @@
+import sys
+sys.path.append("./")
+import gradio as gr
+import torch
+from ip_adapter.utils import BLOCKS as BLOCKS
+import numpy as np
+import random
+from diffusers import (
+    AutoencoderKL,
+    StableDiffusionXLPipeline,
+)
+from ip_adapter import StyleStudio_Adapter
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+dtype = torch.float16 if str(device).__contains__("cuda") else torch.float32
+import os
+os.system("git lfs install")
+os.system("git clone https://huggingface.co/h94/IP-Adapter")
+os.system("mv IP-Adapter/sdxl_models sdxl_models")
+
+from huggingface_hub import hf_hub_download
+
+# hf_hub_download(repo_id="h94/IP-Adapter", filename="sdxl_models/image_encoder", local_dir="./sdxl_models/image_encoder")
+hf_hub_download(repo_id="InstantX/CSGO", filename="csgo_4_32.bin", local_dir="./CSGO/")
+os.system('rm -rf IP-Adapter/models')
+base_model_path = "stabilityai/stable-diffusion-xl-base-1.0"
+image_encoder_path = "sdxl_models/image_encoder"
+csgo_ckpt ='./CSGO/csgo_4_32.bin'
+pretrained_vae_name_or_path ='madebyollin/sdxl-vae-fp16-fix'
+weight_dtype = torch.float16
+
+vae = AutoencoderKL.from_pretrained(pretrained_vae_name_or_path,torch_dtype=torch.float16)
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    base_model_path,
+    torch_dtype=torch.float16,
+    add_watermarker=False,
+    vae=vae
+)
+pipe.enable_vae_tiling()
+
+target_style_blocks = BLOCKS['style']
+
+csgo = StyleStudio_Adapter(
+        pipe, image_encoder_path, csgo_ckpt, device, num_style_tokens=32,
+        target_style_blocks=target_style_blocks,
+        controlnet_adapter=False,
+        style_model_resampler=True,
+
+        fuSAttn=True,
+        end_fusion=20,
+        adainIP=True,
+        )
+
+MAX_SEED = np.iinfo(np.int32).max
+
+
+def get_example():
+    case = [
+        [
+            './assets/style1.jpg',
+            "Text-Driven Style Synthesis",
+            "A red apple",
+            7.0,
+            42,
+            20,
+         ],
+    ]
+    return case
+
+def run_for_examples(style_image_pil, target, prompt, guidance_scale, seed, end_fusion):
+    
+    return create_image(
+        style_image_pil=style_image_pil,
+        prompt=prompt,
+        guidance_scale=7.0,
+        num_inference_steps=50,
+        seed=42,
+        end_fusion=end_fusion,
+        use_SAttn=True,
+        crossModalAdaIN=True,
+    )
+
+def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
+    if randomize_seed:
+        seed = random.randint(0, MAX_SEED)
+    return seed
+
+def create_image(
+                 style_image_pil,
+                 prompt,
+                 guidance_scale,
+                 num_inference_steps,
+                 end_fusion,
+                 crossModalAdaIN,
+                 use_SAttn,
+                 seed,
+                 negative_prompt="text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
+):
+
+    style_image = style_image_pil
+
+    generator = torch.Generator(device).manual_seed(seed)
+    init_latents = torch.randn((1, 4, 128, 128), generator=generator, device="cuda", dtype=torch.float16)
+    num_sample=1
+    if use_SAttn:
+        num_sample=2
+        init_latents = init_latents.repeat(num_sample, 1, 1, 1)
+    with torch.no_grad():
+        images = csgo.generate(pil_style_image=style_image,
+                                prompt=prompt,
+                                negative_prompt=negative_prompt,
+                                height=1024,
+                                width=1024,
+                                guidance_scale=guidance_scale,
+                                num_images_per_prompt=1,
+                                num_samples=num_sample,
+                                num_inference_steps=num_inference_steps,
+                                end_fusion=end_fusion,
+                                cross_modal_adain=crossModalAdaIN,
+                                use_SAttn=use_SAttn,
+                                
+                                generator=generator,
+                                )
+
+    if use_SAttn:
+        return [images[1]]
+    else:
+        return [images[0]]
+
+# Description
+title = r"""
+<h1 align="center">StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements</h1>
+"""
+
+description = r"""
+<b>Official 🤗 Gradio demo</b> for <a href='https://github.com/MingKunLei/StyleStudio' target='_blank'><b>StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements</b></a>.<br>
+How to use:<br>
+1. Upload a style image.
+2. <b>Enter your desired prompt<b>.
+3. Click the <b>Submit</b> button to begin customization.
+4. Share your stylized photo with your friends and enjoy! 😊
+
+Advanced usage:<br>
+1. Click advanced options.
+2. Choose different guidance and steps.
+3. Set the timing for the Teacher Model's participation 
+"""
+
+article = r"""
+---
+📝 **Tips**
+As the value of end_fusion increases, the style gradually diminishes.
+---
+📝 **Citation**
+<br>
+If our work is helpful for your research or applications, please cite us via:
+```bibtex
+
+```
+📧 **Contact**
+<br>
+If you have any questions, please feel free to open an issue or directly reach us out at <b>leimingkun@westlake.edu.cn</b>.
+"""
+
+block = gr.Blocks(css="footer {visibility: hidden}").queue(max_size=10, api_open=False)
+with block:
+    gr.Markdown(title)
+    gr.Markdown(description)
+
+    with gr.Tabs():
+        with gr.Row():
+            with gr.Column():
+                with gr.Row():
+                    with gr.Column():
+                        style_image_pil = gr.Image(label="Style Image", type='pil')
+
+                target = gr.Radio(["Text-Driven Style Synthesis"],
+                                  value="Text-Driven Style Synthesis",
+                                  label="task")
+
+                prompt = gr.Textbox(label="Prompt",
+                                    value="A red apple")
+                
+                neg_prompt = gr.Textbox(label="Negative Prompt",
+                                    value="text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry")
+
+                with gr.Accordion(open=True, label="Advanced Options"):
+
+                    guidance_scale = gr.Slider(minimum=1, maximum=15.0, step=0.01, value=7.0, label="guidance scale")
+                    
+                    num_inference_steps = gr.Slider(minimum=5, maximum=100.0, step=1.0, value=50,
+                                                    label="num inference steps")
+                    
+                    end_fusion = gr.Slider(minimum=0, maximum=num_inference_steps, step=1.0, value=20.0, label="end fusion")
+                    
+                    seed = gr.Slider(minimum=-1000000, maximum=1000000, value=1, step=1, label="Seed Value")
+                    
+                    randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
+                    
+                    crossModalAdaIN = gr.Checkbox(label="Cross Modal AdaIN", value=True)
+                    use_SAttn = gr.Checkbox(label="Teacher Model", value=True)
+
+                generate_button = gr.Button("Generate Image")
+
+            with gr.Column():
+                generated_image = gr.Gallery(label="Generated Image")
+
+        generate_button.click(
+            fn=randomize_seed_fn,
+            inputs=[seed, randomize_seed],
+            outputs=seed,
+            queue=False,
+            api_name=False,
+        ).then(
+            fn=create_image,
+            inputs=[
+                    style_image_pil,
+                    prompt,
+                    guidance_scale,
+                    num_inference_steps,
+                    end_fusion,
+                    crossModalAdaIN,
+                    use_SAttn,
+                    seed,
+                    neg_prompt,],
+            outputs=[generated_image])
+
+    gr.Examples(
+        examples=get_example(),
+        inputs=[style_image_pil, target, prompt, guidance_scale, seed, end_fusion],
+        fn=run_for_examples,
+        outputs=[generated_image],
+        cache_examples=True,
+    )
+
+    gr.Markdown(article)
+
+block.launch()

ip_adapter/__init__.py

@@ -0,0 +1,15 @@
+from .ip_adapter import IPAdapter, IPAdapterPlus, IPAdapterPlusXL, IPAdapterXL, IPAdapterFull,IPAdapterXL_CS,IPAdapter_CS
+from .ip_adapter import CSGO
+from .ip_adapter import StyleStudio_Adapter, StyleStudio_Adapter_exp
+from .ip_adapter import IPAdapterXL_cross_modal
+__all__ = [
+    "IPAdapter",
+    "IPAdapterPlus",
+    "IPAdapterPlusXL",
+    "IPAdapterXL",
+    "CSGO",
+    "StyleStudio_Adapter",
+    "StyleStudio_Adapter_exp",
+    "IPAdapterXL_cross_modal",
+    "IPAdapterFull",
+]

ip_adapter/attention_processor.py

@@ -0,0 +1,1645 @@
+# modified from https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.fft as fft
+import pdb
+
+
+class AttnProcessor(nn.Module):
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+        save_in_unet='down',
+        atten_control=None,
+    ):
+        super().__init__()
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor(nn.Module):
+    r"""
+    Attention processor for IP-Adapater.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,save_in_unet='down', atten_control=None):
+        super().__init__()
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+        self.skip = skip
+
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        if not self.skip:
+            # for ip-adapter
+            ip_key = self.to_k_ip(ip_hidden_states)
+            ip_value = self.to_v_ip(ip_hidden_states)
+
+            ip_key = attn.head_to_batch_dim(ip_key)
+            ip_value = attn.head_to_batch_dim(ip_value)
+
+            ip_attention_probs = attn.get_attention_scores(query, ip_key, None)
+            self.attn_map = ip_attention_probs
+            ip_hidden_states = torch.bmm(ip_attention_probs, ip_value)
+            ip_hidden_states = attn.batch_to_head_dim(ip_hidden_states)
+
+            hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+        save_in_unet='down',
+            atten_control=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IPAttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,save_in_unet='down', atten_control=None):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+        self.skip = skip
+
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if not self.skip:
+            # for ip-adapter
+            ip_key = self.to_k_ip(ip_hidden_states)
+            ip_value = self.to_v_ip(ip_hidden_states)
+
+            ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            ip_hidden_states = F.scaled_dot_product_attention(
+                query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+            with torch.no_grad():
+                self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
+                #print(self.attn_map.shape)
+
+            ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            ip_hidden_states = ip_hidden_states.to(query.dtype)
+
+            hidden_states = hidden_states + self.scale * ip_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class IP_CS_AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, content_scale=1.0,style_scale=1.0, num_content_tokens=4,num_style_tokens=4,
+                 skip=False,content=False, style=False):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.content_scale = content_scale
+        self.style_scale = style_scale
+        self.num_content_tokens = num_content_tokens
+        self.num_style_tokens = num_style_tokens
+        self.skip = skip
+
+        self.content = content
+        self.style = style
+
+        if self.content or self.style:
+            self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_k_ip_content =None
+        self.to_v_ip_content =None
+
+    def set_content_ipa(self,content_scale=1.0):
+
+        self.to_k_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
+        self.to_v_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
+        self.content_scale=content_scale
+        self.content =True
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_content_tokens-self.num_style_tokens
+            encoder_hidden_states, ip_content_hidden_states,ip_style_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:end_pos + self.num_content_tokens, :],
+                encoder_hidden_states[:, end_pos + self.num_content_tokens:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        if self.content is True:
+            exit()
+        if not self.skip and self.content is True:
+            # print('content#####################################################')
+            # for ip-content-adapter
+            if self.to_k_ip_content is None:
+
+                ip_content_key = self.to_k_ip(ip_content_hidden_states)
+                ip_content_value = self.to_v_ip(ip_content_hidden_states)
+            else:
+                ip_content_key = self.to_k_ip_content(ip_content_hidden_states)
+                ip_content_value = self.to_v_ip_content(ip_content_hidden_states)
+
+            ip_content_key = ip_content_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_content_value = ip_content_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            ip_content_hidden_states = F.scaled_dot_product_attention(
+                query, ip_content_key, ip_content_value, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+
+
+            ip_content_hidden_states = ip_content_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            ip_content_hidden_states = ip_content_hidden_states.to(query.dtype)
+
+            hidden_states = hidden_states + self.content_scale * ip_content_hidden_states
+
+        if not self.skip and self.style is True:
+            # for ip-style-adapter
+            ip_style_key = self.to_k_ip(ip_style_hidden_states)
+            ip_style_value = self.to_v_ip(ip_style_hidden_states)
+
+            ip_style_key = ip_style_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_style_value = ip_style_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            ip_style_hidden_states = F.scaled_dot_product_attention(
+                query, ip_style_key, ip_style_value, attn_mask=None, dropout_p=0.0, is_causal=False
+            )
+
+            ip_style_hidden_states = ip_style_hidden_states.transpose(1, 2).reshape(batch_size, -1,
+                                                                                    attn.heads * head_dim)
+            ip_style_hidden_states = ip_style_hidden_states.to(query.dtype)
+
+            hidden_states = hidden_states + self.style_scale * ip_style_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+## for controlnet
+class CNAttnProcessor:
+    r"""
+    Default processor for performing attention-related computations.
+    """
+
+    def __init__(self, num_tokens=4,save_in_unet='down',atten_control=None):
+        self.num_tokens = num_tokens
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+
+    def __call__(self, attn, hidden_states, encoder_hidden_states=None, attention_mask=None, temb=None):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+        attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        query = attn.head_to_batch_dim(query)
+        key = attn.head_to_batch_dim(key)
+        value = attn.head_to_batch_dim(value)
+
+        attention_probs = attn.get_attention_scores(query, key, attention_mask)
+        hidden_states = torch.bmm(attention_probs, value)
+        hidden_states = attn.batch_to_head_dim(hidden_states)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+
+class CNAttnProcessor2_0:
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(self, num_tokens=4, save_in_unet='down', atten_control=None):
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.num_tokens = num_tokens
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+
+    def __call__(
+            self,
+            attn,
+            hidden_states,
+            encoder_hidden_states=None,
+            attention_mask=None,
+            temb=None,
+    ):
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states = encoder_hidden_states[:, :end_pos]  # only use text
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        hidden_states = F.scaled_dot_product_attention(
+            query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+        )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        return hidden_states
+
+class IP_FuAd_AttnProcessor2_0(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, content_scale=1.0,style_scale=1.0, num_content_tokens=4,num_style_tokens=4,
+                 skip=False,content=False, style=False, fuAttn=False, fuIPAttn=False, adainIP=False,
+                 fuScale=0, end_fusion=0, attn_name=None):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.content_scale = content_scale
+        self.style_scale = style_scale
+        self.num_style_tokens = num_style_tokens
+        self.skip = skip
+
+        self.content = content
+        self.style = style
+
+        self.fuAttn = fuAttn
+        self.fuIPAttn = fuIPAttn
+        self.adainIP = adainIP
+        self.fuScale = fuScale
+        self.denoise_step = 0
+        self.end_fusion = end_fusion
+        self.name = attn_name
+
+        if self.content or self.style:
+            self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_k_ip_content =None
+        self.to_v_ip_content =None
+
+    # def set_content_ipa(self,content_scale=1.0):
+
+    #     self.to_k_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
+    #     self.to_v_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
+    #     self.content_scale=content_scale
+    #     self.content =True
+    
+    def reset_denoise_step(self):
+        if self.denoise_step == 50:
+            self.denoise_step = 0
+            # if "up_blocks.0.attentions.1.transformer_blocks.0.attn2" in self.name:
+            #     print("attn2 reset successful")
+    
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        self.denoise_step += 1
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+        
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] -self.num_style_tokens
+            encoder_hidden_states, ip_style_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        # # modified the attnMap of the Stylization Image
+        
+        if self.fuAttn and self.denoise_step <= self.end_fusion:
+            assert query.shape[0] == 4
+            scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+            text_attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
+            text_attn_probs[1] = self.fuScale*text_attn_probs[1] + (1-self.fuScale)*text_attn_probs[0]
+            text_attn_probs[3] = self.fuScale*text_attn_probs[3] + (1-self.fuScale)*text_attn_probs[2]
+            hidden_states = torch.matmul(text_attn_probs, value)
+        else:
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+        
+        raw_hidden_states = hidden_states
+
+        if not self.skip and self.style is True:
+            
+            # for ip-style-adapter
+            ip_style_key = self.to_k_ip(ip_style_hidden_states)
+            ip_style_value = self.to_v_ip(ip_style_hidden_states)
+
+            ip_style_key = ip_style_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_style_value = ip_style_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            if self.fuIPAttn and self.denoise_step <= self.end_fusion:
+                assert query.shape[0] == 4
+                if "down" in self.name:
+                    print("wrong! coding")
+                    exit()
+                scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+                ip_attn_probs = torch.matmul(query, ip_style_key.transpose(-2, -1)) * scale_factor
+                ip_attn_probs = F.softmax(ip_attn_probs, dim=-1)
+                ip_attn_probs[1] = self.fuScale*ip_attn_probs[1] + (1-self.fuScale)*ip_attn_probs[0]
+                ip_attn_probs[3] = self.fuScale*ip_attn_probs[3] + (1-self.fuScale)*ip_attn_probs[2]
+                ip_style_hidden_states = torch.matmul(ip_attn_probs, ip_style_value)
+            else:
+                ip_style_hidden_states = F.scaled_dot_product_attention(
+                    query, ip_style_key, ip_style_value, attn_mask=None, dropout_p=0.0, is_causal=False
+                )
+
+            ip_style_hidden_states = ip_style_hidden_states.transpose(1, 2).reshape(batch_size, -1,
+                                                                                    attn.heads * head_dim)
+            ip_style_hidden_states = ip_style_hidden_states.to(query.dtype)
+
+            if not self.adainIP:
+                hidden_states = hidden_states + self.style_scale * ip_style_hidden_states
+            else:
+                # print("adain")
+                def adain(content, style):
+                    content_mean = content.mean(dim=1, keepdim=True)
+                    content_std = content.std(dim=1, keepdim=True)
+                    style_mean = style.mean(dim=1, keepdim=True)
+                    style_std = style.std(dim=1, keepdim=True)
+                    normalized_content = (content - content_mean) / content_std
+                    stylized_content = normalized_content * style_std + style_mean
+                    return stylized_content
+                hidden_states = adain(content=hidden_states, style=ip_style_hidden_states)
+
+        if hidden_states.shape[0] == 4:
+            hidden_states[0] = raw_hidden_states[0]
+            hidden_states[2] = raw_hidden_states[2]
+        # hidden_states = raw_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        self.reset_denoise_step()
+        return hidden_states
+
+class IP_FuAd_AttnProcessor2_0_exp(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, content_scale=1.0,style_scale=1.0, num_content_tokens=4,num_style_tokens=4,
+                 skip=False,content=False, style=False, fuAttn=False, fuIPAttn=False, adainIP=False,
+                 fuScale=0, end_fusion=0, attn_name=None, save_attn_map=False):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.content_scale = content_scale
+        self.style_scale = style_scale
+        self.num_style_tokens = num_style_tokens
+        self.skip = skip
+
+        self.content = content
+        self.style = style
+
+        self.fuAttn = fuAttn
+        self.fuIPAttn = fuIPAttn
+        self.adainIP = adainIP
+        self.fuScale = fuScale
+        self.denoise_step = 0
+        self.end_fusion = end_fusion
+        self.name = attn_name
+
+        self.save_attn_map = save_attn_map
+
+        if self.content or self.style:
+            self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+            self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_k_ip_content =None
+        self.to_v_ip_content =None
+
+    # def set_content_ipa(self,content_scale=1.0):
+
+    #     self.to_k_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
+    #     self.to_v_ip_content = nn.Linear(self.cross_attention_dim or self.hidden_size, self.hidden_size, bias=False)
+    #     self.content_scale=content_scale
+    #     self.content =True
+    def reset_denoise_step(self):
+        if self.denoise_step == 50:
+            self.denoise_step = 0
+            # if "up_blocks.0.attentions.1.transformer_blocks.0.attn2" in self.name:
+            #     print("attn2 reset successful")
+    
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        self.denoise_step += 1
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+        
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_content_tokens-self.num_style_tokens
+            encoder_hidden_states, ip_style_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        ## attention map
+        if self.save_attn_map:
+            attention_probs = attn.get_attention_scores(attn.head_to_batch_dim(query), attn.head_to_batch_dim(value), attention_mask)
+            if attention_probs is not None:
+                if not hasattr(attn, "attn_map"):
+                    setattr(attn, "attn_map", {})
+                    setattr(attn, "inference_step", 0)
+                else:
+                    attn.inference_step += 1
+                
+                # # maybe we need to save all the timestep
+                # if attn.inference_step in self.attn_map_save_steps:
+                attn.attn_map[attn.inference_step] = attention_probs.clone().cpu().detach()
+                # attn.attn_map[attn.inference_step] = attention_probs.detach()
+                ## end of attention map
+            else:
+                print(f"{attn} didn't get the attention probs")
+        
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        # # modified the attnMap of the Stylization Image
+        
+        if self.fuAttn and self.denoise_step <= self.end_fusion:
+            assert query.shape[0] == 4
+            scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+            text_attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
+            text_attn_probs[1] = self.fuScale*text_attn_probs[1] + (1-self.fuScale)*text_attn_probs[0]
+            text_attn_probs[3] = self.fuScale*text_attn_probs[3] + (1-self.fuScale)*text_attn_probs[2]
+            hidden_states = torch.matmul(text_attn_probs, value)
+        else:
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+        
+        raw_hidden_states = hidden_states
+
+        if not self.skip and self.style is True:
+            
+            # for ip-style-adapter
+            ip_style_key = self.to_k_ip(ip_style_hidden_states)
+            ip_style_value = self.to_v_ip(ip_style_hidden_states)
+
+            ip_style_key = ip_style_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_style_value = ip_style_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            if self.fuIPAttn and self.denoise_step <= self.end_fusion:
+                assert query.shape[0] == 4
+                if "down" in self.name:
+                    print("wrong! coding")
+                    exit()
+                scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+                ip_attn_probs = torch.matmul(query, ip_style_key.transpose(-2, -1)) * scale_factor
+                ip_attn_probs = F.softmax(ip_attn_probs, dim=-1)
+                ip_attn_probs[1] = self.fuScale*ip_attn_probs[1] + (1-self.fuScale)*ip_attn_probs[0]
+                ip_attn_probs[3] = self.fuScale*ip_attn_probs[3] + (1-self.fuScale)*ip_attn_probs[2]
+                ip_style_hidden_states = torch.matmul(ip_attn_probs, ip_style_value)
+            else:
+                ip_style_hidden_states = F.scaled_dot_product_attention(
+                    query, ip_style_key, ip_style_value, attn_mask=None, dropout_p=0.0, is_causal=False
+                )
+
+            ip_style_hidden_states = ip_style_hidden_states.transpose(1, 2).reshape(batch_size, -1,
+                                                                                    attn.heads * head_dim)
+            ip_style_hidden_states = ip_style_hidden_states.to(query.dtype)
+
+            # if self.adainIP and self.denoise_step >= self.start_adain:
+            if self.adainIP:
+                # print("adain")
+                # if self.denoise_step == 1 and "up_blocks.1.attentions.2.transformer_blocks.1" in self.name:
+                    # print("adain")
+                def adain(content, style):
+                    content_mean = content.mean(dim=1, keepdim=True)
+                    content_std = content.std(dim=1, keepdim=True)
+                    print("exp code")
+                    pdb.set_trace()
+                    style_mean = style.mean(dim=1, keepdim=True)
+                    style_std = style.std(dim=1, keepdim=True)
+                    normalized_content = (content - content_mean) / content_std
+                    stylized_content = normalized_content * style_std + style_mean
+                    return stylized_content
+                pdb.set_trace()
+                hidden_states = adain(content=hidden_states, style=ip_style_hidden_states)
+            else:
+                hidden_states = hidden_states + self.style_scale * ip_style_hidden_states
+
+        if hidden_states.shape[0] == 4:
+            hidden_states[0] = raw_hidden_states[0]
+            hidden_states[2] = raw_hidden_states[2]
+        # hidden_states = raw_hidden_states
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        self.reset_denoise_step()
+        return hidden_states
+
+class AttnProcessor2_0_hijack(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+        save_in_unet='down',
+        atten_control=None,
+        fuSAttn=False,
+        fuScale=0,
+        end_fusion=0,
+        attn_name=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+        
+        self.fuSAttn = fuSAttn
+        self.fuScale = fuScale
+        self.denoise_step = 0
+        self.end_fusion = end_fusion
+        self.name = attn_name
+
+    def reset_denoise_step(self):
+        if self.denoise_step == 50:
+            self.denoise_step = 0
+            # if "up_blocks.0.attentions.1.transformer_blocks.0.attn1" in self.name:
+            #     print("attn1 reset successful")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        self.denoise_step += 1
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        if self.fuSAttn and self.denoise_step <= self.end_fusion:
+            assert query.shape[0] == 4
+            if "up_blocks.1.attentions.2.transformer_blocks.1" in self.name and self.denoise_step == self.end_fusion:
+                print("now: ", self.denoise_step, "end now:", self.end_fusion, "scale: ", self.fuScale)
+                # pdb.set_trace()
+            scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+            attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
+            attn_probs[1] = self.fuScale*attn_probs[1] + (1-self.fuScale)*attn_probs[0]
+            attn_probs[3] = self.fuScale*attn_probs[3] + (1-self.fuScale)*attn_probs[2]
+            hidden_states = torch.matmul(attn_probs, value)
+        else:
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        if self.denoise_step == 50:
+            self.reset_denoise_step()
+        return hidden_states
+    
+class AttnProcessor2_0_exp(torch.nn.Module):
+    r"""
+    Processor for implementing scaled dot-product attention (enabled by default if you're using PyTorch 2.0).
+    """
+
+    def __init__(
+        self,
+        hidden_size=None,
+        cross_attention_dim=None,
+        save_in_unet='down',
+        atten_control=None,
+        fuSAttn=False,
+        fuScale=0,
+        end_fusion=0,
+        attn_name=None,
+    ):
+        super().__init__()
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+        self.atten_control = atten_control
+        self.save_in_unet = save_in_unet
+        
+        self.fuSAttn = fuSAttn
+        self.fuScale = fuScale
+        self.denoise_step = 0
+        self.end_fusion = end_fusion
+        self.name = attn_name
+
+    def reset_denoise_step(self):
+        if self.denoise_step == 50:
+            self.denoise_step = 0
+            # if "up_blocks.0.attentions.1.transformer_blocks.0.attn1" in self.name:
+            #     print("attn1 reset successful")
+
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        self.denoise_step += 1
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        elif attn.norm_cross:
+            encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        if self.fuSAttn and self.denoise_step <= self.end_fusion:
+            assert query.shape[0] == 4
+            if "up_blocks.1.attentions.2.transformer_blocks.1" in self.name and self.denoise_step == self.end_fusion:
+                print("now: ", self.denoise_step, "end now:", self.end_fusion, "scale: ", self.fuScale)
+                # pdb.set_trace()
+            scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+            attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
+            
+            attn_probs[1] = self.fuScale*attn_probs[1] + (1-self.fuScale)*attn_probs[0]
+            attn_probs[3] = self.fuScale*attn_probs[3] + (1-self.fuScale)*attn_probs[2]
+            print("exp code")
+            pdb.set_trace()
+            def adain(content, style):
+                content_mean = content.mean(dim=1, keepdim=True)
+                content_std = content.std(dim=1, keepdim=True)
+                style_mean = style.mean(dim=1, keepdim=True)
+                style_std = style.std(dim=1, keepdim=True)
+                normalized_content = (content - content_mean) / content_std
+                stylized_content = normalized_content * style_std + style_mean
+                return stylized_content
+            value[1] = adain(content=value[0], style=value[1])
+            value[3] = adain(content=value[2], style=value[3])
+            hidden_states = torch.matmul(attn_probs, value)
+        else:
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        self.reset_denoise_step()
+        return hidden_states
+    
+class IPAttnProcessor2_0_cross_modal(torch.nn.Module):
+    r"""
+    Attention processor for IP-Adapater for PyTorch 2.0.
+    Args:
+        hidden_size (`int`):
+            The hidden size of the attention layer.
+        cross_attention_dim (`int`):
+            The number of channels in the `encoder_hidden_states`.
+        scale (`float`, defaults to 1.0):
+            the weight scale of image prompt.
+        num_tokens (`int`, defaults to 4 when do ip_adapter_plus it should be 16):
+            The context length of the image features.
+    """
+
+    def __init__(self, hidden_size, cross_attention_dim=None, scale=1.0, num_tokens=4, skip=False,
+                 fuAttn=False, fuIPAttn=False, adainIP=False, end_fusion=0, fuScale=0, attn_name=None):
+        super().__init__()
+
+        if not hasattr(F, "scaled_dot_product_attention"):
+            raise ImportError("AttnProcessor2_0 requires PyTorch 2.0, to use it, please upgrade PyTorch to 2.0.")
+
+        self.hidden_size = hidden_size
+        self.cross_attention_dim = cross_attention_dim
+        self.scale = scale
+        self.num_tokens = num_tokens
+        self.skip = skip
+
+        self.fuAttn = fuAttn
+        self.fuIPAttn = fuIPAttn
+        self.adainIP = adainIP
+        self.denoise_step = fuScale
+        self.end_fusion = end_fusion
+        self.fuScale = fuScale
+        self.name = attn_name
+
+        self.to_k_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+        self.to_v_ip = nn.Linear(cross_attention_dim or hidden_size, hidden_size, bias=False)
+
+    def reset_denoise_step(self):
+        if self.denoise_step == 50:
+            self.denoise_step = 0
+            # if "up_blocks.0.attentions.1.transformer_blocks.0.attn2" in self.name:
+            #     print("attn2 reset successful")
+    
+    def __call__(
+        self,
+        attn,
+        hidden_states,
+        encoder_hidden_states=None,
+        attention_mask=None,
+        temb=None,
+    ):
+        self.denoise_step += 1
+        residual = hidden_states
+
+        if attn.spatial_norm is not None:
+            hidden_states = attn.spatial_norm(hidden_states, temb)
+
+        input_ndim = hidden_states.ndim
+
+        if input_ndim == 4:
+            batch_size, channel, height, width = hidden_states.shape
+            hidden_states = hidden_states.view(batch_size, channel, height * width).transpose(1, 2)
+
+        batch_size, sequence_length, _ = (
+            hidden_states.shape if encoder_hidden_states is None else encoder_hidden_states.shape
+        )
+
+        if attention_mask is not None:
+            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
+            # scaled_dot_product_attention expects attention_mask shape to be
+            # (batch, heads, source_length, target_length)
+            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
+
+        if attn.group_norm is not None:
+            hidden_states = attn.group_norm(hidden_states.transpose(1, 2)).transpose(1, 2)
+
+        query = attn.to_q(hidden_states)
+
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+        else:
+            # get encoder_hidden_states, ip_hidden_states
+            end_pos = encoder_hidden_states.shape[1] - self.num_tokens
+            encoder_hidden_states, ip_hidden_states = (
+                encoder_hidden_states[:, :end_pos, :],
+                encoder_hidden_states[:, end_pos:, :],
+            )
+            if attn.norm_cross:
+                encoder_hidden_states = attn.norm_encoder_hidden_states(encoder_hidden_states)
+
+        key = attn.to_k(encoder_hidden_states)
+        value = attn.to_v(encoder_hidden_states)
+
+        inner_dim = key.shape[-1]
+        head_dim = inner_dim // attn.heads
+
+        query = query.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        key = key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+        value = value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+        # the output of sdp = (batch, num_heads, seq_len, head_dim)
+        # TODO: add support for attn.scale when we move to Torch 2.1
+        if self.fuAttn and self.denoise_step <= self.end_fusion:
+            assert query.shape[0] == 4
+            if "up_blocks.1.attentions.2.transformer_blocks.1" in self.name and self.denoise_step == self.end_fusion:
+                print("fuAttn")
+                print("now: ", self.denoise_step, "end now:", self.end_fusion, "scale: ", self.fuScale)
+            scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+            text_attn_probs = (torch.matmul(query, key.transpose(-2, -1)) * scale_factor).softmax(dim=-1)
+            text_attn_probs[1] = self.fuScale*text_attn_probs[1] + (1-self.fuScale)*text_attn_probs[0]
+            text_attn_probs[3] = self.fuScale*text_attn_probs[3] + (1-self.fuScale)*text_attn_probs[2]
+            hidden_states = torch.matmul(text_attn_probs, value)
+        else:
+            hidden_states = F.scaled_dot_product_attention(
+                query, key, value, attn_mask=attention_mask, dropout_p=0.0, is_causal=False
+            )
+
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+        hidden_states = hidden_states.to(query.dtype)
+
+        raw_hidden_states = hidden_states
+
+        if not self.skip:
+            # for ip-adapter
+            ip_key = self.to_k_ip(ip_hidden_states)
+            ip_value = self.to_v_ip(ip_hidden_states)
+
+            ip_key = ip_key.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+            ip_value = ip_value.view(batch_size, -1, attn.heads, head_dim).transpose(1, 2)
+
+            # the output of sdp = (batch, num_heads, seq_len, head_dim)
+            # TODO: add support for attn.scale when we move to Torch 2.1
+            if self.fuIPAttn and self.denoise_step <= self.end_fusion:
+                assert query.shape[0] == 4
+                print("fuIPAttn")
+                if "down" in self.name:
+                    print("wrong! coding")
+                    exit()
+                scale_factor = 1 / math.sqrt(torch.tensor(head_dim, dtype=query.dtype))
+                ip_attn_probs = torch.matmul(query, ip_key.transpose(-2, -1)) * scale_factor
+                ip_attn_probs = F.softmax(ip_attn_probs, dim=-1)
+                ip_attn_probs[1] = self.fuScale*ip_attn_probs[1] + (1-self.fuScale)*ip_attn_probs[0]
+                ip_attn_probs[3] = self.fuScale*ip_attn_probs[3] + (1-self.fuScale)*ip_attn_probs[2]
+                ip_hidden_states = torch.matmul(ip_attn_probs, ip_value)
+            else:
+                ip_hidden_states = F.scaled_dot_product_attention(
+                    query, ip_key, ip_value, attn_mask=None, dropout_p=0.0, is_causal=False
+                )
+
+            with torch.no_grad():
+                self.attn_map = query @ ip_key.transpose(-2, -1).softmax(dim=-1)
+                #print(self.attn_map.shape)
+
+            ip_hidden_states = ip_hidden_states.transpose(1, 2).reshape(batch_size, -1, attn.heads * head_dim)
+            ip_hidden_states = ip_hidden_states.to(query.dtype)
+            
+            if self.adainIP:
+                def adain(content, style):
+                    # 计算内容特征的均值和标准差
+                    content_mean = content.mean(dim=1, keepdim=True)
+                    content_std = content.std(dim=1, keepdim=True)
+                    # 计算风格特征的均值和标准差
+                    style_mean = style.mean(dim=1, keepdim=True)
+                    style_std = style.std(dim=1, keepdim=True)
+                    # 归一化内容特征并应用风格特征的均值和方差
+                    normalized_content = (content - content_mean) / content_std
+                    stylized_content = normalized_content * style_std + style_mean
+                    return stylized_content
+                hidden_states = adain(content=hidden_states, style=ip_hidden_states)
+            else:
+                hidden_states = hidden_states + self.scale * ip_hidden_states
+        
+        if hidden_states.shape[0] == 4:
+            hidden_states[0] = raw_hidden_states[0]
+            hidden_states[2] = raw_hidden_states[2]
+        
+        # linear proj
+        hidden_states = attn.to_out[0](hidden_states)
+        # dropout
+        hidden_states = attn.to_out[1](hidden_states)
+
+        if input_ndim == 4:
+            hidden_states = hidden_states.transpose(-1, -2).reshape(batch_size, channel, height, width)
+
+        if attn.residual_connection:
+            hidden_states = hidden_states + residual
+
+        hidden_states = hidden_states / attn.rescale_output_factor
+
+        if self.denoise_step == 50:
+            self.reset_denoise_step()
+
+        return hidden_states

ip_adapter/ip_adapter.py

@@ -0,0 +1,1757 @@
+import torch.nn.functional as F
+import os
+from typing import List
+
+import torch
+from diffusers import StableDiffusionPipeline
+from diffusers.pipelines.controlnet import MultiControlNetModel
+from PIL import Image
+from safetensors import safe_open
+from transformers import CLIPImageProcessor, CLIPVisionModelWithProjection
+from torchvision import transforms
+from .utils import is_torch2_available, get_generator
+
+if is_torch2_available():
+    from .attention_processor import (
+        AttnProcessor2_0 as AttnProcessor,
+    )
+    from .attention_processor import (
+        CNAttnProcessor2_0 as CNAttnProcessor,
+    )
+    from .attention_processor import (
+        IPAttnProcessor2_0 as IPAttnProcessor,
+    )
+    from .attention_processor import IP_CS_AttnProcessor2_0 as IP_CS_AttnProcessor
+    from .attention_processor import IP_FuAd_AttnProcessor2_0 as IP_FuAd_AttnProcessor
+    from .attention_processor import IP_FuAd_AttnProcessor2_0_exp as IP_FuAd_AttnProcessor_exp
+    from .attention_processor import AttnProcessor2_0_exp as AttnProcessor_exp
+    from .attention_processor import AttnProcessor2_0_hijack as AttnProcessor_hijack
+    from .attention_processor import IPAttnProcessor2_0_cross_modal as IPAttnProcessor_cross_modal
+else:
+    from .attention_processor import AttnProcessor, CNAttnProcessor, IPAttnProcessor
+
+from .resampler import Resampler
+
+from transformers import AutoImageProcessor, AutoModel
+
+
+class ImageProjModel(torch.nn.Module):
+    """Projection Model"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
+        super().__init__()
+
+        self.generator = None
+        self.cross_attention_dim = cross_attention_dim
+        self.clip_extra_context_tokens = clip_extra_context_tokens
+        # print(clip_embeddings_dim, self.clip_extra_context_tokens, cross_attention_dim)
+        self.proj = torch.nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
+        self.norm = torch.nn.LayerNorm(cross_attention_dim)
+
+    def forward(self, image_embeds):
+        embeds = image_embeds
+        clip_extra_context_tokens = self.proj(embeds).reshape(
+            -1, self.clip_extra_context_tokens, self.cross_attention_dim
+        )
+        clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
+        return clip_extra_context_tokens
+
+
+class MLPProjModel(torch.nn.Module):
+    """SD model with image prompt"""
+
+    def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024):
+        super().__init__()
+
+        self.proj = torch.nn.Sequential(
+            torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
+            torch.nn.GELU(),
+            torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
+            torch.nn.LayerNorm(cross_attention_dim)
+        )
+
+    def forward(self, image_embeds):
+        clip_extra_context_tokens = self.proj(image_embeds)
+        return clip_extra_context_tokens
+
+
+class IPAdapter:
+    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4, target_blocks=["block"]):
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_tokens = num_tokens
+        self.target_blocks = target_blocks
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter()
+
+        # load image encoder
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
+            self.device, dtype=torch.float16
+        )
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.image_proj_model = self.init_proj()
+
+        self.load_ip_adapter()
+
+    def init_proj(self):
+        image_proj_model = ImageProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.projection_dim,
+            clip_extra_context_tokens=self.num_tokens,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor()
+            else:
+                selected = False
+                for block_name in self.target_blocks:
+                    if block_name in name:
+                        selected = True
+                        break
+                if selected:
+                    attn_procs[name] = IPAttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=self.num_tokens,
+                    ).to(self.device, dtype=torch.float16)
+                else:
+                    attn_procs[name] = IPAttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=self.num_tokens,
+                        skip=True
+                    ).to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                for controlnet in self.pipe.controlnet.nets:
+                    controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
+            else:
+                self.pipe.controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
+
+    def load_ip_adapter(self):
+        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
+            state_dict = {"image_proj": {}, "ip_adapter": {}}
+            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    if key.startswith("image_proj."):
+                        state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+        else:
+            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
+        self.image_proj_model.load_state_dict(state_dict["image_proj"])
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None):
+        if pil_image is not None:
+            if isinstance(pil_image, Image.Image):
+                pil_image = [pil_image]
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+        else:
+            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+
+        if content_prompt_embeds is not None:
+            clip_image_embeds = clip_image_embeds - content_prompt_embeds
+
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def set_scale(self, scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IPAttnProcessor):
+                attn_processor.scale = scale
+
+    def generate(
+            self,
+            pil_image=None,
+            clip_image_embeds=None,
+            prompt=None,
+            negative_prompt=None,
+            scale=1.0,
+            num_samples=4,
+            seed=None,
+            guidance_scale=7.5,
+            num_inference_steps=30,
+            neg_content_emb=None,
+            **kwargs,
+    ):
+        self.set_scale(scale)
+
+        if pil_image is not None:
+            num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+        else:
+            num_prompts = clip_image_embeds.size(0)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(
+            pil_image=pil_image, clip_image_embeds=clip_image_embeds, content_prompt_embeds=neg_content_emb
+        )
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
+                prompt,
+                device=self.device,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds_, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds_, uncond_image_prompt_embeds], dim=1)
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapter_CS:
+    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_content_tokens=4,
+                 num_style_tokens=4,
+                 target_content_blocks=["block"], target_style_blocks=["block"], content_image_encoder_path=None,
+                 controlnet_adapter=False,
+                 controlnet_target_content_blocks=None,
+                 controlnet_target_style_blocks=None,
+                 content_model_resampler=False,
+                 style_model_resampler=False,
+                ):
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_content_tokens = num_content_tokens
+        self.num_style_tokens = num_style_tokens
+        self.content_target_blocks = target_content_blocks
+        self.style_target_blocks = target_style_blocks
+
+        self.content_model_resampler = content_model_resampler
+        self.style_model_resampler = style_model_resampler
+
+        self.controlnet_adapter = controlnet_adapter
+        self.controlnet_target_content_blocks = controlnet_target_content_blocks
+        self.controlnet_target_style_blocks = controlnet_target_style_blocks
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter()
+        self.content_image_encoder_path = content_image_encoder_path
+
+
+        # load image encoder
+        if content_image_encoder_path is not None:
+            self.content_image_encoder = AutoModel.from_pretrained(content_image_encoder_path).to(self.device,
+                                                                                                  dtype=torch.float16)
+            self.content_image_processor = AutoImageProcessor.from_pretrained(content_image_encoder_path)
+        else:
+            self.content_image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
+                self.device, dtype=torch.float16
+            )
+            self.content_image_processor = CLIPImageProcessor()
+        # model.requires_grad_(False)
+
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
+            self.device, dtype=torch.float16
+        )
+        # if self.use_CSD is not None:
+        #     self.style_image_encoder = CSD_CLIP("vit_large", "default",self.use_CSD+"/ViT-L-14.pt")
+        #     model_path = self.use_CSD+"/checkpoint.pth"
+        #     checkpoint = torch.load(model_path, map_location="cpu")
+        #     state_dict = convert_state_dict(checkpoint['model_state_dict'])
+        #     self.style_image_encoder.load_state_dict(state_dict, strict=False)
+        #
+        #     normalize = transforms.Normalize((0.48145466, 0.4578275, 0.40821073), (0.26862954, 0.26130258, 0.27577711))
+        #     self.style_preprocess = transforms.Compose([
+        #         transforms.Resize(size=224, interpolation=Func.InterpolationMode.BICUBIC),
+        #         transforms.CenterCrop(224),
+        #         transforms.ToTensor(),
+        #         normalize,
+        #     ])
+
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.content_image_proj_model = self.init_proj(self.num_content_tokens, content_or_style_='content',
+                                                       model_resampler=self.content_model_resampler)
+        self.style_image_proj_model = self.init_proj(self.num_style_tokens, content_or_style_='style',
+                                                     model_resampler=self.style_model_resampler)
+
+        self.load_ip_adapter()
+
+    def init_proj(self, num_tokens, content_or_style_='content', model_resampler=False):
+
+        # print('@@@@',self.pipe.unet.config.cross_attention_dim,self.image_encoder.config.projection_dim)
+        if content_or_style_ == 'content' and self.content_image_encoder_path is not None:
+            image_proj_model = ImageProjModel(
+                cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+                clip_embeddings_dim=self.content_image_encoder.config.projection_dim,
+                clip_extra_context_tokens=num_tokens,
+            ).to(self.device, dtype=torch.float16)
+            return image_proj_model
+
+        image_proj_model = ImageProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.projection_dim,
+            clip_extra_context_tokens=num_tokens,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor()
+            else:
+                # layername_id += 1
+                selected = False
+                for block_name in self.style_target_blocks:
+                    if block_name in name:
+                        selected = True
+                        # print(name)
+                        attn_procs[name] = IP_CS_AttnProcessor(
+                            hidden_size=hidden_size,
+                            cross_attention_dim=cross_attention_dim,
+                            style_scale=1.0,
+                            style=True,
+                            num_content_tokens=self.num_content_tokens,
+                            num_style_tokens=self.num_style_tokens,
+                        )
+                for block_name in self.content_target_blocks:
+                    if block_name in name:
+                        # selected = True
+                        if selected is False:
+                            attn_procs[name] = IP_CS_AttnProcessor(
+                                hidden_size=hidden_size,
+                                cross_attention_dim=cross_attention_dim,
+                                content_scale=1.0,
+                                content=True,
+                                num_content_tokens=self.num_content_tokens,
+                                num_style_tokens=self.num_style_tokens,
+                            )
+                        else:
+                            attn_procs[name].set_content_ipa(content_scale=1.0)
+                            # attn_procs[name].content=True
+
+                if selected is False:
+                    attn_procs[name] = IP_CS_AttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        num_content_tokens=self.num_content_tokens,
+                        num_style_tokens=self.num_style_tokens,
+                        skip=True,
+                    )
+
+                attn_procs[name].to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if self.controlnet_adapter is False:
+                if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                    for controlnet in self.pipe.controlnet.nets:
+                        controlnet.set_attn_processor(CNAttnProcessor(
+                            num_tokens=self.num_content_tokens + self.num_style_tokens))
+                else:
+                    self.pipe.controlnet.set_attn_processor(CNAttnProcessor(
+                        num_tokens=self.num_content_tokens + self.num_style_tokens))
+
+            else:
+                controlnet_attn_procs = {}
+                controlnet_style_target_blocks = self.controlnet_target_style_blocks
+                controlnet_content_target_blocks = self.controlnet_target_content_blocks
+                for name in self.pipe.controlnet.attn_processors.keys():
+                    # print(name)
+                    cross_attention_dim = None if name.endswith(
+                        "attn1.processor") else self.pipe.controlnet.config.cross_attention_dim
+                    if name.startswith("mid_block"):
+                        hidden_size = self.pipe.controlnet.config.block_out_channels[-1]
+                    elif name.startswith("up_blocks"):
+                        block_id = int(name[len("up_blocks.")])
+                        hidden_size = list(reversed(self.pipe.controlnet.config.block_out_channels))[block_id]
+                    elif name.startswith("down_blocks"):
+                        block_id = int(name[len("down_blocks.")])
+                        hidden_size = self.pipe.controlnet.config.block_out_channels[block_id]
+                    if cross_attention_dim is None:
+                        # layername_id += 1
+                        controlnet_attn_procs[name] = AttnProcessor()
+
+                    else:
+                        # layername_id += 1
+                        selected = False
+                        for block_name in controlnet_style_target_blocks:
+                            if block_name in name:
+                                selected = True
+                                # print(name)
+                                controlnet_attn_procs[name] = IP_CS_AttnProcessor(
+                                    hidden_size=hidden_size,
+                                    cross_attention_dim=cross_attention_dim,
+                                    style_scale=1.0,
+                                    style=True,
+                                    num_content_tokens=self.num_content_tokens,
+                                    num_style_tokens=self.num_style_tokens,
+                                )
+
+                        for block_name in controlnet_content_target_blocks:
+                            if block_name in name:
+                                if selected is False:
+                                    controlnet_attn_procs[name] = IP_CS_AttnProcessor(
+                                        hidden_size=hidden_size,
+                                        cross_attention_dim=cross_attention_dim,
+                                        content_scale=1.0,
+                                        content=True,
+                                        num_content_tokens=self.num_content_tokens,
+                                        num_style_tokens=self.num_style_tokens,
+                                    )
+
+                                    selected = True
+                                elif selected is True:
+                                    controlnet_attn_procs[name].set_content_ipa(content_scale=1.0)
+
+                                # if args.content_image_encoder_type !='dinov2':
+                                #     weights = {
+                                #         "to_k_ip.weight": state_dict["ip_adapter"][str(layername_id) + ".to_k_ip.weight"],
+                                #         "to_v_ip.weight": state_dict["ip_adapter"][str(layername_id) + ".to_v_ip.weight"],
+                                #     }
+                                #     attn_procs[name].load_state_dict(weights)
+                        if selected is False:
+                            controlnet_attn_procs[name] = IP_CS_AttnProcessor(
+                                hidden_size=hidden_size,
+                                cross_attention_dim=cross_attention_dim,
+                                num_content_tokens=self.num_content_tokens,
+                                num_style_tokens=self.num_style_tokens,
+                                skip=True,
+                            )
+                        controlnet_attn_procs[name].to(self.device, dtype=torch.float16)
+                        # layer_name = name.split(".processor")[0]
+                        # # print(state_dict["ip_adapter"].keys())
+                        # weights = {
+                        #     "to_k_ip.weight": state_dict["ip_adapter"][str(layername_id) + ".to_k_ip.weight"],
+                        #     "to_v_ip.weight": state_dict["ip_adapter"][str(layername_id) + ".to_v_ip.weight"],
+                        # }
+                        # attn_procs[name].load_state_dict(weights)
+                self.pipe.controlnet.set_attn_processor(controlnet_attn_procs)
+
+    def load_ip_adapter(self):
+        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
+            state_dict = {"content_image_proj": {}, "style_image_proj": {}, "ip_adapter": {}}
+            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    if key.startswith("content_image_proj."):
+                        state_dict["content_image_proj"][key.replace("content_image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("style_image_proj."):
+                        state_dict["style_image_proj"][key.replace("style_image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+        else:
+            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
+        self.content_image_proj_model.load_state_dict(state_dict["content_image_proj"])
+        self.style_image_proj_model.load_state_dict(state_dict["style_image_proj"])
+
+        if 'conv_in_unet_sd' in state_dict.keys():
+            self.pipe.unet.conv_in.load_state_dict(state_dict["conv_in_unet_sd"], strict=True)
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
+
+        if self.controlnet_adapter is True:
+            print('loading controlnet_adapter')
+            self.pipe.controlnet.load_state_dict(state_dict["controlnet_adapter_modules"], strict=False)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None,
+                         content_or_style_=''):
+        # if pil_image is not None:
+        #     if isinstance(pil_image, Image.Image):
+        #         pil_image = [pil_image]
+        #     clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        #     clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+        # else:
+        #     clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+
+        # if content_prompt_embeds is not None:
+        #     clip_image_embeds = clip_image_embeds - content_prompt_embeds
+
+        if content_or_style_ == 'content':
+            if pil_image is not None:
+                if isinstance(pil_image, Image.Image):
+                    pil_image = [pil_image]
+                if self.content_image_proj_model is not None:
+                    clip_image = self.content_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                    clip_image_embeds = self.content_image_encoder(
+                        clip_image.to(self.device, dtype=torch.float16)).image_embeds
+                else:
+                    clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                    clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+            else:
+                clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+
+            image_prompt_embeds = self.content_image_proj_model(clip_image_embeds)
+            uncond_image_prompt_embeds = self.content_image_proj_model(torch.zeros_like(clip_image_embeds))
+            return image_prompt_embeds, uncond_image_prompt_embeds
+        if content_or_style_ == 'style':
+            if pil_image is not None:
+                if self.use_CSD is not None:
+                    clip_image = self.style_preprocess(pil_image).unsqueeze(0).to(self.device, dtype=torch.float32)
+                    clip_image_embeds = self.style_image_encoder(clip_image)
+                else:
+                    if isinstance(pil_image, Image.Image):
+                        pil_image = [pil_image]
+                    clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                    clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+
+
+            else:
+                clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+            image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+            uncond_image_prompt_embeds = self.style_image_proj_model(torch.zeros_like(clip_image_embeds))
+            return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def set_scale(self, content_scale, style_scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IP_CS_AttnProcessor):
+                if attn_processor.content is True:
+                    attn_processor.content_scale = content_scale
+
+                if attn_processor.style is True:
+                    attn_processor.style_scale = style_scale
+                    # print('style_scale:',style_scale)
+        if self.controlnet_adapter is not None:
+            for attn_processor in self.pipe.controlnet.attn_processors.values():
+
+                if isinstance(attn_processor, IP_CS_AttnProcessor):
+                    if attn_processor.content is True:
+                        attn_processor.content_scale = content_scale
+                        # print(content_scale)
+
+                    if attn_processor.style is True:
+                        attn_processor.style_scale = style_scale
+
+    def generate(
+            self,
+            pil_content_image=None,
+            pil_style_image=None,
+            clip_content_image_embeds=None,
+            clip_style_image_embeds=None,
+            prompt=None,
+            negative_prompt=None,
+            content_scale=1.0,
+            style_scale=1.0,
+            num_samples=4,
+            seed=None,
+            guidance_scale=7.5,
+            num_inference_steps=30,
+            neg_content_emb=None,
+            **kwargs,
+    ):
+        self.set_scale(content_scale, style_scale)
+
+        if pil_content_image is not None:
+            num_prompts = 1 if isinstance(pil_content_image, Image.Image) else len(pil_content_image)
+        else:
+            num_prompts = clip_content_image_embeds.size(0)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        content_image_prompt_embeds, uncond_content_image_prompt_embeds = self.get_image_embeds(
+            pil_image=pil_content_image, clip_image_embeds=clip_content_image_embeds
+        )
+        style_image_prompt_embeds, uncond_style_image_prompt_embeds = self.get_image_embeds(
+            pil_image=pil_style_image, clip_image_embeds=clip_style_image_embeds
+        )
+
+        bs_embed, seq_len, _ = content_image_prompt_embeds.shape
+        content_image_prompt_embeds = content_image_prompt_embeds.repeat(1, num_samples, 1)
+        content_image_prompt_embeds = content_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_content_image_prompt_embeds = uncond_content_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_content_image_prompt_embeds = uncond_content_image_prompt_embeds.view(bs_embed * num_samples, seq_len,
+                                                                                     -1)
+
+        bs_style_embed, seq_style_len, _ = content_image_prompt_embeds.shape
+        style_image_prompt_embeds = style_image_prompt_embeds.repeat(1, num_samples, 1)
+        style_image_prompt_embeds = style_image_prompt_embeds.view(bs_embed * num_samples, seq_style_len, -1)
+        uncond_style_image_prompt_embeds = uncond_style_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_style_image_prompt_embeds = uncond_style_image_prompt_embeds.view(bs_embed * num_samples, seq_style_len,
+                                                                                 -1)
+
+        with torch.inference_mode():
+            prompt_embeds_, negative_prompt_embeds_ = self.pipe.encode_prompt(
+                prompt,
+                device=self.device,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds_, content_image_prompt_embeds, style_image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds_,
+                                                uncond_content_image_prompt_embeds, uncond_style_image_prompt_embeds],
+                                               dim=1)
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            guidance_scale=guidance_scale,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapterXL_CS(IPAdapter_CS):
+    """SDXL"""
+
+    def generate(
+            self,
+            pil_content_image,
+            pil_style_image,
+            prompt=None,
+            negative_prompt=None,
+            content_scale=1.0,
+            style_scale=1.0,
+            num_samples=4,
+            seed=None,
+            content_image_embeds=None,
+            style_image_embeds=None,
+            num_inference_steps=30,
+            neg_content_emb=None,
+            neg_content_prompt=None,
+            neg_content_scale=1.0,
+
+            **kwargs,
+    ):
+        self.set_scale(content_scale, style_scale)
+
+        num_prompts = 1 if isinstance(pil_content_image, Image.Image) else len(pil_content_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        content_image_prompt_embeds, uncond_content_image_prompt_embeds = self.get_image_embeds(pil_content_image,
+                                                                                                content_image_embeds,
+                                                                                                content_or_style_='content')
+
+
+
+        style_image_prompt_embeds, uncond_style_image_prompt_embeds = self.get_image_embeds(pil_style_image,
+                                                                                            style_image_embeds,
+                                                                                            content_or_style_='style')
+
+    
+        bs_embed, seq_len, _ = content_image_prompt_embeds.shape
+
+        content_image_prompt_embeds = content_image_prompt_embeds.repeat(1, num_samples, 1)
+        content_image_prompt_embeds = content_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        uncond_content_image_prompt_embeds = uncond_content_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_content_image_prompt_embeds = uncond_content_image_prompt_embeds.view(bs_embed * num_samples, seq_len,
+                                                                                     -1)
+        bs_style_embed, seq_style_len, _ = style_image_prompt_embeds.shape
+        style_image_prompt_embeds = style_image_prompt_embeds.repeat(1, num_samples, 1)
+        style_image_prompt_embeds = style_image_prompt_embeds.view(bs_embed * num_samples, seq_style_len, -1)
+        uncond_style_image_prompt_embeds = uncond_style_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_style_image_prompt_embeds = uncond_style_image_prompt_embeds.view(bs_embed * num_samples, seq_style_len,
+                                                                                 -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, content_image_prompt_embeds, style_image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds,
+                                                uncond_content_image_prompt_embeds, uncond_style_image_prompt_embeds],
+                                               dim=1)
+
+        # self.generator = get_generator(seed, self.device)
+        # latents = torch.randn((1, 4, 128, 128), generator=self.generator, device="cuda", dtype=torch.float16).to("cuda")
+        # latents = latents.repeat(2, 1, 1, 1)
+        # print(latents.shape)
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            # generator=self.generator,
+            **kwargs,
+        ).images
+        return images
+
+
+class CSGO(IPAdapterXL_CS):
+    """SDXL"""
+
+    def init_proj(self, num_tokens, content_or_style_='content', model_resampler=False):
+        if content_or_style_ == 'content':
+            if model_resampler:
+                image_proj_model = Resampler(
+                    dim=self.pipe.unet.config.cross_attention_dim,
+                    depth=4,
+                    dim_head=64,
+                    heads=12,
+                    num_queries=num_tokens,
+                    embedding_dim=self.content_image_encoder.config.hidden_size,
+                    output_dim=self.pipe.unet.config.cross_attention_dim,
+                    ff_mult=4,
+                ).to(self.device, dtype=torch.float16)
+            else:
+                image_proj_model = ImageProjModel(
+                    cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+                    clip_embeddings_dim=self.image_encoder.config.projection_dim,
+                    clip_extra_context_tokens=num_tokens,
+                ).to(self.device, dtype=torch.float16)
+        if content_or_style_ == 'style':
+            if model_resampler:
+                image_proj_model = Resampler(
+                    dim=self.pipe.unet.config.cross_attention_dim,
+                    depth=4,
+                    dim_head=64,
+                    heads=12,
+                    num_queries=num_tokens,
+                    embedding_dim=self.content_image_encoder.config.hidden_size,
+                    output_dim=self.pipe.unet.config.cross_attention_dim,
+                    ff_mult=4,
+                ).to(self.device, dtype=torch.float16)
+            else:
+                image_proj_model = ImageProjModel(
+                    cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+                    clip_embeddings_dim=self.image_encoder.config.projection_dim,
+                    clip_extra_context_tokens=num_tokens,
+                ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None, content_or_style_=''):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        if content_or_style_ == 'style':
+            
+            if self.style_model_resampler:
+                clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16),
+                                                       output_hidden_states=True).hidden_states[-2]
+                image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+                uncond_image_prompt_embeds = self.style_image_proj_model(torch.zeros_like(clip_image_embeds))
+            else:
+
+
+                clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+                image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+                uncond_image_prompt_embeds = self.style_image_proj_model(torch.zeros_like(clip_image_embeds))
+            return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+        else:
+
+            if self.content_image_encoder_path is not None:
+                clip_image = self.content_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                outputs = self.content_image_encoder(clip_image.to(self.device, dtype=torch.float16),
+                                                     output_hidden_states=True)
+                clip_image_embeds = outputs.last_hidden_state
+                image_prompt_embeds = self.content_image_proj_model(clip_image_embeds)
+
+                # uncond_clip_image_embeds = self.image_encoder(
+                #     torch.zeros_like(clip_image), output_hidden_states=True
+                # ).last_hidden_state
+                uncond_image_prompt_embeds = self.content_image_proj_model(torch.zeros_like(clip_image_embeds))
+                return image_prompt_embeds, uncond_image_prompt_embeds
+
+            else:
+                if self.content_model_resampler:
+
+                    clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+
+                    clip_image = clip_image.to(self.device, dtype=torch.float16)
+                    clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+                    # clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+                    image_prompt_embeds = self.content_image_proj_model(clip_image_embeds)
+                    # uncond_clip_image_embeds = self.image_encoder(
+                    #             torch.zeros_like(clip_image), output_hidden_states=True
+                    #         ).hidden_states[-2]
+                    uncond_image_prompt_embeds = self.content_image_proj_model(torch.zeros_like(clip_image_embeds))
+                else:
+                    clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+                    clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+                    image_prompt_embeds = self.content_image_proj_model(clip_image_embeds)
+                    uncond_image_prompt_embeds = self.content_image_proj_model(torch.zeros_like(clip_image_embeds))
+
+                return image_prompt_embeds, uncond_image_prompt_embeds
+
+        #     # clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        #     clip_image = clip_image.to(self.device, dtype=torch.float16)
+        #     clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        #     image_prompt_embeds = self.content_image_proj_model(clip_image_embeds)
+        #     uncond_clip_image_embeds = self.image_encoder(
+        #         torch.zeros_like(clip_image), output_hidden_states=True
+        #     ).hidden_states[-2]
+        #     uncond_image_prompt_embeds = self.content_image_proj_model(uncond_clip_image_embeds)
+        # return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class StyleStudio_Adapter(CSGO):
+    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device,
+                 num_style_tokens=4,
+                 target_style_blocks=["block"], 
+                 controlnet_adapter=False,
+                 controlnet_target_content_blocks=None,
+                 controlnet_target_style_blocks=None,
+                 style_model_resampler=False,
+                 fuAttn=False,
+                 fuSAttn=False,
+                 fuIPAttn=False,
+                 fuScale=0,
+                 adainIP=False,
+                 end_fusion=0,
+                 save_attn_map=False,
+                ):
+        self.fuAttn = fuAttn
+        self.fuSAttn = fuSAttn
+        self.fuIPAttn = fuIPAttn
+        self.adainIP = adainIP
+        self.fuScale = fuScale
+        # if self.adainIP:
+        #     print("use the cross modal adain")
+        if self.fuSAttn:
+            print(f"hijack Self AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
+        if self.fuAttn:
+            print(f"hijack Cross AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
+        if self.fuIPAttn:
+            print(f"hijack IP AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
+        self.end_fusion = end_fusion
+        self.save_attn_map = save_attn_map
+
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_style_tokens = num_style_tokens
+        self.style_target_blocks = target_style_blocks
+
+        self.style_model_resampler = style_model_resampler
+
+        self.controlnet_adapter = controlnet_adapter
+        self.controlnet_target_content_blocks = controlnet_target_content_blocks
+        self.controlnet_target_style_blocks = controlnet_target_style_blocks
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter()
+
+
+        # load image encoder
+        # model.requires_grad_(False)
+
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
+            self.device, dtype=torch.float16
+        )
+
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.style_image_proj_model = self.init_proj(self.num_style_tokens, content_or_style_='style',
+                                                     model_resampler=self.style_model_resampler)
+        self.load_ip_adapter()
+
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor_hijack(
+                                        fuSAttn=self.fuSAttn,
+                                        fuScale=self.fuScale,
+                                        end_fusion=self.end_fusion,
+                                        attn_name=name)
+            else:
+                # layername_id += 1
+                selected = False
+                for block_name in self.style_target_blocks:
+                    if block_name in name:
+                        selected = True
+                        # print(name)
+                        attn_procs[name] = IP_FuAd_AttnProcessor(
+                            hidden_size=hidden_size,
+                            cross_attention_dim=cross_attention_dim,
+                            style_scale=1.0,
+                            style=True,
+                            num_style_tokens=self.num_style_tokens,
+                            fuAttn=self.fuAttn,
+                            fuIPAttn=self.fuIPAttn,
+                            adainIP=self.adainIP,
+                            fuScale=self.fuScale,
+                            end_fusion=self.end_fusion,
+                            attn_name=name,
+                        )
+                if selected is False:
+                    attn_procs[name] = IP_FuAd_AttnProcessor(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        num_style_tokens=self.num_style_tokens,
+                        skip=True,
+                        fuAttn=self.fuAttn,
+
+                        fuIPAttn=self.fuIPAttn,
+                        adainIP=self.adainIP,
+                        fuScale=self.fuScale,
+                        end_fusion=self.end_fusion,
+                        attn_name=name,
+                    )
+
+                attn_procs[name].to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if self.controlnet_adapter is False:
+                if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                    for controlnet in self.pipe.controlnet.nets:
+                        controlnet.set_attn_processor(CNAttnProcessor(
+                            num_tokens=self.num_content_tokens + self.num_style_tokens))
+                else:
+                    self.pipe.controlnet.set_attn_processor(CNAttnProcessor(
+                        num_tokens=self.num_content_tokens + self.num_style_tokens))
+
+    def load_ip_adapter(self):
+        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
+            state_dict = {"content_image_proj": {}, "style_image_proj": {}, "ip_adapter": {}}
+            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    if key.startswith("content_image_proj."):
+                        state_dict["content_image_proj"][key.replace("content_image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("style_image_proj."):
+                        state_dict["style_image_proj"][key.replace("style_image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+        else:
+            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
+        self.style_image_proj_model.load_state_dict(state_dict["style_image_proj"])
+
+        if 'conv_in_unet_sd' in state_dict.keys():
+            self.pipe.unet.conv_in.load_state_dict(state_dict["conv_in_unet_sd"], strict=True)
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
+    
+    def set_scale(self, style_scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IP_FuAd_AttnProcessor):
+                if attn_processor.style is True:
+                    attn_processor.style_scale = style_scale
+                    # print('style_scale:',style_scale)
+    
+    def init_proj(self, num_tokens, content_or_style_='content', model_resampler=False):
+        if content_or_style_ == 'content':
+            if model_resampler:
+                image_proj_model = Resampler(
+                    dim=self.pipe.unet.config.cross_attention_dim,
+                    depth=4,
+                    dim_head=64,
+                    heads=12,
+                    num_queries=num_tokens,
+                    embedding_dim=self.image_encoder.config.hidden_size,
+                    output_dim=self.pipe.unet.config.cross_attention_dim,
+                    ff_mult=4,
+                ).to(self.device, dtype=torch.float16)
+            else:
+                image_proj_model = ImageProjModel(
+                    cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+                    clip_embeddings_dim=self.image_encoder.config.projection_dim,
+                    clip_extra_context_tokens=num_tokens,
+                ).to(self.device, dtype=torch.float16)
+        if content_or_style_ == 'style':
+            if model_resampler:
+                image_proj_model = Resampler(
+                    dim=self.pipe.unet.config.cross_attention_dim,
+                    depth=4,
+                    dim_head=64,
+                    heads=12,
+                    num_queries=num_tokens,
+                    embedding_dim=self.image_encoder.config.hidden_size,
+                    output_dim=self.pipe.unet.config.cross_attention_dim,
+                    ff_mult=4,
+                ).to(self.device, dtype=torch.float16)
+            else:
+                image_proj_model = ImageProjModel(
+                    cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+                    clip_embeddings_dim=self.image_encoder.config.projection_dim,
+                    clip_extra_context_tokens=num_tokens,
+                ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        if self.style_model_resampler:
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16),
+                                                    output_hidden_states=True).hidden_states[-2]
+            image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+            uncond_image_prompt_embeds = self.style_image_proj_model(torch.zeros_like(clip_image_embeds))
+        else:
+
+
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+            image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+            uncond_image_prompt_embeds = self.style_image_proj_model(torch.zeros_like(clip_image_embeds))
+        return image_prompt_embeds, uncond_image_prompt_embeds
+        
+    @torch.inference_mode()
+    def get_neg_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+
+        if self.style_model_resampler:
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16),
+                                                    output_hidden_states=True).hidden_states[-2]     
+            neg_image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+        else:
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+            neg_image_prompt_embeds = self.style_image_proj_model(clip_image_embeds)
+        return neg_image_prompt_embeds
+    
+    def set_endFusion(self, end_T):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, AttnProcessor_hijack):
+                attn_processor.end_fusion = end_T
+    
+    def set_SAttn(self, use_SAttn):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, AttnProcessor_hijack):
+                attn_processor.fuSAttn = use_SAttn
+    
+    def set_adain(self, use_CMA):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IP_FuAd_AttnProcessor):
+                attn_processor.adainIP = use_CMA
+
+    def generate(
+            self,
+            pil_style_image,
+
+            neg_pil_style_image=None,
+            
+            prompt=None,
+            negative_prompt=None,
+            num_samples=2,
+            style_image_embeds=None,
+            num_inference_steps=30,
+            end_fusion=20,
+            cross_modal_adain=True,
+            use_SAttn=True,
+            **kwargs,
+    ):
+        
+        self.set_endFusion(end_T = end_fusion)
+        self.set_adain(use_CMA=cross_modal_adain)
+        self.set_SAttn(use_SAttn=use_SAttn)
+        
+        # self.set_scale(style_scale=style_scale)
+        num_prompts = 1 if isinstance(pil_style_image, Image.Image) else len(pil_style_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        style_image_prompt_embeds, uncond_style_image_prompt_embeds = self.get_image_embeds(
+            pil_style_image,
+            style_image_embeds,
+        )
+
+        if neg_pil_style_image is not None:
+            print("using neg style image")
+            neg_style_image_prompt_embeds = self.get_neg_image_embeds(neg_pil_style_image,
+                                                                        style_image_embeds,)
+            cos_sim_neg = F.cosine_similarity(style_image_prompt_embeds, neg_style_image_prompt_embeds.squeeze(0).unsqueeze(1), dim=-1)
+            cos_sim_uncond = F.cosine_similarity(style_image_prompt_embeds, uncond_style_image_prompt_embeds.squeeze(0).unsqueeze(1), dim=-1)
+            print(f"neg cos sim is: {cos_sim_neg.diagonal()}")
+            print(f"uncond cos sim is: {cos_sim_uncond.diagonal()}")
+            uncond_style_image_prompt_embeds = neg_style_image_prompt_embeds
+
+        bs_embed, seq_style_len, _ = style_image_prompt_embeds.shape
+        style_image_prompt_embeds = style_image_prompt_embeds.repeat(1, num_samples, 1)
+        style_image_prompt_embeds = style_image_prompt_embeds.view(bs_embed * num_samples, seq_style_len, -1)
+        uncond_style_image_prompt_embeds = uncond_style_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_style_image_prompt_embeds = uncond_style_image_prompt_embeds.view(bs_embed * num_samples, seq_style_len,
+                                                                                 -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, style_image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds,
+                                                uncond_style_image_prompt_embeds],
+                                               dim=1)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            **kwargs,
+        ).images
+        return images
+
+# StyleStudio_Adapter experiment code
+class StyleStudio_Adapter_exp(StyleStudio_Adapter):
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor_exp(
+                                        fuSAttn=self.fuSAttn,
+                                        fuScale=self.fuScale,
+                                        end_fusion=self.end_fusion,
+                                        attn_name=name)
+            else:
+                # layername_id += 1
+                selected = False
+                for block_name in self.style_target_blocks:
+                    if block_name in name:
+                        selected = True
+                        # print(name)
+                        #  将所有的StyleBlock中的都改为FuAdAttn
+                        attn_procs[name] = IP_FuAd_AttnProcessor_exp(
+                            hidden_size=hidden_size,
+                            cross_attention_dim=cross_attention_dim,
+                            style_scale=1.0,
+                            style=True,
+                            num_content_tokens=self.num_content_tokens,
+                            num_style_tokens=self.num_style_tokens,
+                            fuAttn=self.fuAttn,
+                            fuIPAttn=self.fuIPAttn,
+                            adainIP=self.adainIP,
+                            fuScale=self.fuScale,
+                            end_fusion=self.end_fusion,
+                            attn_name=name,
+                            save_attn_map=self.save_attn_map,
+                        )
+                # 没有CSGO中关于Content Control的需求 因此就将这个处理Content tokens Cross Attention 删除
+                # 并且这里应该是CSGO代码中 有问题的部分 不论如何这里都会被之后的重置
+                # 并且在CSGO的设计里Content Block和Style Block是没有子集的
+                # selected False表明不是Style Block 关键是 Skip = True
+                if selected is False:
+                    attn_procs[name] = IP_FuAd_AttnProcessor_exp(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        num_content_tokens=self.num_content_tokens,
+                        num_style_tokens=self.num_style_tokens,
+                        skip=True,
+                        fuAttn=self.fuAttn,
+                        fuIPAttn=self.fuIPAttn,
+                        adainIP=self.adainIP,
+                        fuScale=self.fuScale,
+                        end_fusion=self.end_fusion,
+                        attn_name=name,
+                        save_attn_map=self.save_attn_map,
+                    )
+                    # attn_procs[name] = IP_FuAd_AttnProcessor_exp(
+                    #     hidden_size=hidden_size,
+                    #     cross_attention_dim=cross_attention_dim,
+                    #     num_content_tokens=self.num_content_tokens,
+                    #     num_style_tokens=self.num_style_tokens,
+                    #     skip=True,
+                    #     fuAttn=self.fuAttn,
+                    #     fuIPAttn=self.fuIPAttn,
+                    # )
+
+                attn_procs[name].to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if self.controlnet_adapter is False:
+                if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                    for controlnet in self.pipe.controlnet.nets:
+                        controlnet.set_attn_processor(CNAttnProcessor(
+                            num_tokens=self.num_content_tokens + self.num_style_tokens))
+                else:
+                    self.pipe.controlnet.set_attn_processor(CNAttnProcessor(
+                        num_tokens=self.num_content_tokens + self.num_style_tokens))
+            # 因为我们的代码中没有controlnet需要将Style 注入 这并不是一个I2I的任务
+            # 因此我们将原本CSGO中和ControlNet中注入Style的部分给删除了
+
+class IPAdapterXL(IPAdapter):
+    """SDXL"""
+
+    def generate(
+            self,
+            pil_image,
+            prompt=None,
+            negative_prompt=None,
+            scale=1.0,
+            num_samples=4,
+            seed=None,
+            num_inference_steps=30,
+            neg_content_emb=None,
+            neg_content_prompt=None,
+            neg_content_scale=1.0,
+            **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        if neg_content_emb is None:
+            if neg_content_prompt is not None:
+                with torch.inference_mode():
+                    (
+                        prompt_embeds_,  # torch.Size([1, 77, 2048])
+                        negative_prompt_embeds_,
+                        pooled_prompt_embeds_,  # torch.Size([1, 1280])
+                        negative_pooled_prompt_embeds_,
+                    ) = self.pipe.encode_prompt(
+                        neg_content_prompt,
+                        num_images_per_prompt=num_samples,
+                        do_classifier_free_guidance=True,
+                        negative_prompt=negative_prompt,
+                    )
+                    pooled_prompt_embeds_ *= neg_content_scale
+            else:
+                pooled_prompt_embeds_ = neg_content_emb
+        else:
+            pooled_prompt_embeds_ = None
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image,
+                                                                                content_prompt_embeds=pooled_prompt_embeds_)
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
+
+        self.generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=self.generator,
+            **kwargs,
+        ).images
+
+        return images
+
+
+class IPAdapterXL_cross_modal(IPAdapterXL):
+    def __init__(self, sd_pipe, image_encoder_path, ip_ckpt, device, num_tokens=4, 
+                 target_blocks=["block"],
+                 fuAttn=False,
+                 fuSAttn=False,
+                 fuIPAttn=False,
+                 fuScale=0,
+                 adainIP=False,
+                 end_fusion=0,
+                 save_attn_map=False,):
+        self.fuAttn = fuAttn
+        self.fuSAttn = fuSAttn
+        self.fuIPAttn = fuIPAttn
+        self.adainIP = adainIP
+        self.fuScale = fuScale
+        if self.fuSAttn:
+            print(f"hijack Self AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
+        if self.fuAttn:
+            print(f"hijack Cross AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
+        if self.fuIPAttn:
+            print(f"hijack IP AttnMap in {end_fusion} steps", "fuScale is: ", fuScale)
+        self.end_fusion = end_fusion
+        self.save_attn_map = save_attn_map
+
+        self.device = device
+        self.image_encoder_path = image_encoder_path
+        self.ip_ckpt = ip_ckpt
+        self.num_tokens = num_tokens
+        self.target_blocks = target_blocks
+
+        self.pipe = sd_pipe.to(self.device)
+        self.set_ip_adapter()
+
+        # load image encoder
+        self.image_encoder = CLIPVisionModelWithProjection.from_pretrained(self.image_encoder_path).to(
+            self.device, dtype=torch.float16
+        )
+        self.clip_image_processor = CLIPImageProcessor()
+        # image proj model
+        self.image_proj_model = self.init_proj()
+
+        self.load_ip_adapter()
+
+    def init_proj(self):
+        image_proj_model = ImageProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.projection_dim,
+            clip_extra_context_tokens=self.num_tokens,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+    
+    def set_ip_adapter(self):
+        unet = self.pipe.unet
+        attn_procs = {}
+        for name in unet.attn_processors.keys():
+            cross_attention_dim = None if name.endswith("attn1.processor") else unet.config.cross_attention_dim
+            if name.startswith("mid_block"):
+                hidden_size = unet.config.block_out_channels[-1]
+            elif name.startswith("up_blocks"):
+                block_id = int(name[len("up_blocks.")])
+                hidden_size = list(reversed(unet.config.block_out_channels))[block_id]
+            elif name.startswith("down_blocks"):
+                block_id = int(name[len("down_blocks.")])
+                hidden_size = unet.config.block_out_channels[block_id]
+            if cross_attention_dim is None:
+                attn_procs[name] = AttnProcessor_hijack(
+                                        fuSAttn=self.fuSAttn,
+                                        fuScale=self.fuScale,
+                                        end_fusion=self.end_fusion,
+                                        attn_name=name) # Self Attention
+            else: # Cross Attention
+                selected = False
+                for block_name in self.target_blocks:
+                    if block_name in name:
+                        selected = True
+                        break
+                if selected:
+                    attn_procs[name] = IPAttnProcessor_cross_modal(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=self.num_tokens,
+                        fuAttn=self.fuAttn,
+                        fuIPAttn=self.fuIPAttn,
+                        adainIP=self.adainIP,
+                        fuScale=self.fuScale,
+                        end_fusion=self.end_fusion,
+                        attn_name=name,
+                    ).to(self.device, dtype=torch.float16)
+                else:
+                    attn_procs[name] = IPAttnProcessor_cross_modal(
+                        hidden_size=hidden_size,
+                        cross_attention_dim=cross_attention_dim,
+                        scale=1.0,
+                        num_tokens=self.num_tokens,
+                        skip=True,
+                        fuAttn=self.fuAttn,
+                        fuIPAttn=self.fuIPAttn,
+                        adainIP=self.adainIP,
+                        fuScale=self.fuScale,
+                        end_fusion=self.end_fusion,
+                        attn_name=name,
+                    ).to(self.device, dtype=torch.float16)
+        unet.set_attn_processor(attn_procs)
+        if hasattr(self.pipe, "controlnet"):
+            if isinstance(self.pipe.controlnet, MultiControlNetModel):
+                for controlnet in self.pipe.controlnet.nets:
+                    controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
+            else:
+                self.pipe.controlnet.set_attn_processor(CNAttnProcessor(num_tokens=self.num_tokens))
+
+    def load_ip_adapter(self):
+        if os.path.splitext(self.ip_ckpt)[-1] == ".safetensors":
+            state_dict = {"image_proj": {}, "ip_adapter": {}}
+            with safe_open(self.ip_ckpt, framework="pt", device="cpu") as f:
+                for key in f.keys():
+                    if key.startswith("image_proj."):
+                        state_dict["image_proj"][key.replace("image_proj.", "")] = f.get_tensor(key)
+                    elif key.startswith("ip_adapter."):
+                        state_dict["ip_adapter"][key.replace("ip_adapter.", "")] = f.get_tensor(key)
+        else:
+            state_dict = torch.load(self.ip_ckpt, map_location="cpu")
+        self.image_proj_model.load_state_dict(state_dict["image_proj"])
+        ip_layers = torch.nn.ModuleList(self.pipe.unet.attn_processors.values())
+        ip_layers.load_state_dict(state_dict["ip_adapter"], strict=False)
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None):
+        if pil_image is not None:
+            if isinstance(pil_image, Image.Image):
+                pil_image = [pil_image]
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+        else:
+            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+        
+        if content_prompt_embeds is not None:
+            clip_image_embeds = clip_image_embeds - content_prompt_embeds
+
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_image_prompt_embeds = self.image_proj_model(torch.zeros_like(clip_image_embeds))
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def set_scale(self, scale):
+        for attn_processor in self.pipe.unet.attn_processors.values():
+            if isinstance(attn_processor, IPAttnProcessor_cross_modal):
+                attn_processor.scale = scale
+
+    @torch.inference_mode()
+    def get_neg_image_embeds(self, pil_image=None, clip_image_embeds=None, content_prompt_embeds=None):
+        if pil_image is not None:
+            if isinstance(pil_image, Image.Image):
+                pil_image = [pil_image]
+            clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+            clip_image_embeds = self.image_encoder(clip_image.to(self.device, dtype=torch.float16)).image_embeds
+        else:
+            clip_image_embeds = clip_image_embeds.to(self.device, dtype=torch.float16)
+        
+        if content_prompt_embeds is not None:
+            clip_image_embeds = clip_image_embeds - content_prompt_embeds
+
+        neg_image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        return neg_image_prompt_embeds
+    
+    def generate(
+        self,
+        pil_image,
+        neg_pil_image=None,
+        prompt=None,
+        negative_prompt=None,
+        scale=1.0,
+        num_samples=4,
+        seed=None,
+        num_inference_steps=30,
+        neg_content_emb=None,
+        neg_content_prompt=None,
+        neg_content_scale=1.0,
+        **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+        
+        if neg_content_emb is None:
+            if neg_content_prompt is not None:
+                with torch.inference_mode():
+                    (
+                        prompt_embeds_, # torch.Size([1, 77, 2048])
+                        negative_prompt_embeds_,
+                        pooled_prompt_embeds_, # torch.Size([1, 1280])
+                        negative_pooled_prompt_embeds_,
+                    ) = self.pipe.encode_prompt(
+                        neg_content_prompt,
+                        num_images_per_prompt=num_samples,
+                        do_classifier_free_guidance=True,
+                        negative_prompt=negative_prompt,
+                    )
+                    pooled_prompt_embeds_ *= neg_content_scale
+            else:
+                pooled_prompt_embeds_ = neg_content_emb
+        else:
+            pooled_prompt_embeds_ = None
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image, content_prompt_embeds=pooled_prompt_embeds_)
+        
+        if neg_pil_image is not None:
+            neg_image_prompt_embeds = self.get_neg_image_embeds(neg_pil_image)
+            cos_sim_neg = F.cosine_similarity(image_prompt_embeds, neg_image_prompt_embeds.squeeze(0).unsqueeze(1), dim=-1)
+            cos_sim_uncond = F.cosine_similarity(image_prompt_embeds, uncond_image_prompt_embeds.squeeze(0).unsqueeze(1), dim=-1)
+            print(f"neg cos sim is: {cos_sim_neg.diagonal()}")
+            print(f"uncond cos sim is: {cos_sim_uncond.diagonal()}")
+            uncond_image_prompt_embeds = neg_image_prompt_embeds
+        
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
+
+        # self.generator = get_generator(seed, self.device)
+        
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            # generator=self.generator,
+            **kwargs,
+        ).images
+
+        return images
+    
+
+class IPAdapterPlus(IPAdapter):
+    """IP-Adapter with fine-grained features"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=self.pipe.unet.config.cross_attention_dim,
+            depth=4,
+            dim_head=64,
+            heads=12,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image=None, clip_image_embeds=None):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+
+class IPAdapterFull(IPAdapterPlus):
+    """IP-Adapter with full features"""
+
+    def init_proj(self):
+        image_proj_model = MLPProjModel(
+            cross_attention_dim=self.pipe.unet.config.cross_attention_dim,
+            clip_embeddings_dim=self.image_encoder.config.hidden_size,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+
+class IPAdapterPlusXL(IPAdapter):
+    """SDXL"""
+
+    def init_proj(self):
+        image_proj_model = Resampler(
+            dim=1280,
+            depth=4,
+            dim_head=64,
+            heads=20,
+            num_queries=self.num_tokens,
+            embedding_dim=self.image_encoder.config.hidden_size,
+            output_dim=self.pipe.unet.config.cross_attention_dim,
+            ff_mult=4,
+        ).to(self.device, dtype=torch.float16)
+        return image_proj_model
+
+    @torch.inference_mode()
+    def get_image_embeds(self, pil_image):
+        if isinstance(pil_image, Image.Image):
+            pil_image = [pil_image]
+        clip_image = self.clip_image_processor(images=pil_image, return_tensors="pt").pixel_values
+        clip_image = clip_image.to(self.device, dtype=torch.float16)
+        clip_image_embeds = self.image_encoder(clip_image, output_hidden_states=True).hidden_states[-2]
+        image_prompt_embeds = self.image_proj_model(clip_image_embeds)
+        uncond_clip_image_embeds = self.image_encoder(
+            torch.zeros_like(clip_image), output_hidden_states=True
+        ).hidden_states[-2]
+        uncond_image_prompt_embeds = self.image_proj_model(uncond_clip_image_embeds)
+        return image_prompt_embeds, uncond_image_prompt_embeds
+
+    def generate(
+            self,
+            pil_image,
+            prompt=None,
+            negative_prompt=None,
+            scale=1.0,
+            num_samples=4,
+            seed=None,
+            num_inference_steps=30,
+            **kwargs,
+    ):
+        self.set_scale(scale)
+
+        num_prompts = 1 if isinstance(pil_image, Image.Image) else len(pil_image)
+
+        if prompt is None:
+            prompt = "best quality, high quality"
+        if negative_prompt is None:
+            negative_prompt = "monochrome, lowres, bad anatomy, worst quality, low quality"
+
+        if not isinstance(prompt, List):
+            prompt = [prompt] * num_prompts
+        if not isinstance(negative_prompt, List):
+            negative_prompt = [negative_prompt] * num_prompts
+
+        image_prompt_embeds, uncond_image_prompt_embeds = self.get_image_embeds(pil_image)
+        bs_embed, seq_len, _ = image_prompt_embeds.shape
+        image_prompt_embeds = image_prompt_embeds.repeat(1, num_samples, 1)
+        image_prompt_embeds = image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.repeat(1, num_samples, 1)
+        uncond_image_prompt_embeds = uncond_image_prompt_embeds.view(bs_embed * num_samples, seq_len, -1)
+
+        with torch.inference_mode():
+            (
+                prompt_embeds,
+                negative_prompt_embeds,
+                pooled_prompt_embeds,
+                negative_pooled_prompt_embeds,
+            ) = self.pipe.encode_prompt(
+                prompt,
+                num_images_per_prompt=num_samples,
+                do_classifier_free_guidance=True,
+                negative_prompt=negative_prompt,
+            )
+            prompt_embeds = torch.cat([prompt_embeds, image_prompt_embeds], dim=1)
+            negative_prompt_embeds = torch.cat([negative_prompt_embeds, uncond_image_prompt_embeds], dim=1)
+
+        generator = get_generator(seed, self.device)
+
+        images = self.pipe(
+            prompt_embeds=prompt_embeds,
+            negative_prompt_embeds=negative_prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
+            num_inference_steps=num_inference_steps,
+            generator=generator,
+            **kwargs,
+        ).images
+
+        return images

ip_adapter/resampler.py

@@ -0,0 +1,158 @@
+# modified from https://github.com/mlfoundations/open_flamingo/blob/main/open_flamingo/src/helpers.py
+# and https://github.com/lucidrains/imagen-pytorch/blob/main/imagen_pytorch/imagen_pytorch.py
+
+import math
+
+import torch
+import torch.nn as nn
+from einops import rearrange
+from einops.layers.torch import Rearrange
+
+
+# FFN
+def FeedForward(dim, mult=4):
+    inner_dim = int(dim * mult)
+    return nn.Sequential(
+        nn.LayerNorm(dim),
+        nn.Linear(dim, inner_dim, bias=False),
+        nn.GELU(),
+        nn.Linear(inner_dim, dim, bias=False),
+    )
+
+
+def reshape_tensor(x, heads):
+    bs, length, width = x.shape
+    # (bs, length, width) --> (bs, length, n_heads, dim_per_head)
+    x = x.view(bs, length, heads, -1)
+    # (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
+    x = x.transpose(1, 2)
+    # (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
+    x = x.reshape(bs, heads, length, -1)
+    return x
+
+
+class PerceiverAttention(nn.Module):
+    def __init__(self, *, dim, dim_head=64, heads=8):
+        super().__init__()
+        self.scale = dim_head**-0.5
+        self.dim_head = dim_head
+        self.heads = heads
+        inner_dim = dim_head * heads
+
+        self.norm1 = nn.LayerNorm(dim)
+        self.norm2 = nn.LayerNorm(dim)
+
+        self.to_q = nn.Linear(dim, inner_dim, bias=False)
+        self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
+        self.to_out = nn.Linear(inner_dim, dim, bias=False)
+
+    def forward(self, x, latents):
+        """
+        Args:
+            x (torch.Tensor): image features
+                shape (b, n1, D)
+            latent (torch.Tensor): latent features
+                shape (b, n2, D)
+        """
+        x = self.norm1(x)
+        latents = self.norm2(latents)
+
+        b, l, _ = latents.shape
+
+        q = self.to_q(latents)
+        kv_input = torch.cat((x, latents), dim=-2)
+        k, v = self.to_kv(kv_input).chunk(2, dim=-1)
+
+        q = reshape_tensor(q, self.heads)
+        k = reshape_tensor(k, self.heads)
+        v = reshape_tensor(v, self.heads)
+
+        # attention
+        scale = 1 / math.sqrt(math.sqrt(self.dim_head))
+        weight = (q * scale) @ (k * scale).transpose(-2, -1)  # More stable with f16 than dividing afterwards
+        weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
+        out = weight @ v
+
+        out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
+
+        return self.to_out(out)
+
+
+class Resampler(nn.Module):
+    def __init__(
+        self,
+        dim=1024,
+        depth=8,
+        dim_head=64,
+        heads=16,
+        num_queries=8,
+        embedding_dim=768,
+        output_dim=1024,
+        ff_mult=4,
+        max_seq_len: int = 257,  # CLIP tokens + CLS token
+        apply_pos_emb: bool = False,
+        num_latents_mean_pooled: int = 0,  # number of latents derived from mean pooled representation of the sequence
+    ):
+        super().__init__()
+        self.pos_emb = nn.Embedding(max_seq_len, embedding_dim) if apply_pos_emb else None
+
+        self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
+
+        self.proj_in = nn.Linear(embedding_dim, dim)
+
+        self.proj_out = nn.Linear(dim, output_dim)
+        self.norm_out = nn.LayerNorm(output_dim)
+
+        self.to_latents_from_mean_pooled_seq = (
+            nn.Sequential(
+                nn.LayerNorm(dim),
+                nn.Linear(dim, dim * num_latents_mean_pooled),
+                Rearrange("b (n d) -> b n d", n=num_latents_mean_pooled),
+            )
+            if num_latents_mean_pooled > 0
+            else None
+        )
+
+        self.layers = nn.ModuleList([])
+        for _ in range(depth):
+            self.layers.append(
+                nn.ModuleList(
+                    [
+                        PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
+                        FeedForward(dim=dim, mult=ff_mult),
+                    ]
+                )
+            )
+
+    def forward(self, x):
+        if self.pos_emb is not None:
+            n, device = x.shape[1], x.device
+            pos_emb = self.pos_emb(torch.arange(n, device=device))
+            x = x + pos_emb
+
+        latents = self.latents.repeat(x.size(0), 1, 1)
+
+        x = self.proj_in(x)
+
+        if self.to_latents_from_mean_pooled_seq:
+            meanpooled_seq = masked_mean(x, dim=1, mask=torch.ones(x.shape[:2], device=x.device, dtype=torch.bool))
+            meanpooled_latents = self.to_latents_from_mean_pooled_seq(meanpooled_seq)
+            latents = torch.cat((meanpooled_latents, latents), dim=-2)
+
+        for attn, ff in self.layers:
+            latents = attn(x, latents) + latents
+            latents = ff(latents) + latents
+
+        latents = self.proj_out(latents)
+        return self.norm_out(latents)
+
+
+def masked_mean(t, *, dim, mask=None):
+    if mask is None:
+        return t.mean(dim=dim)
+
+    denom = mask.sum(dim=dim, keepdim=True)
+    mask = rearrange(mask, "b n -> b n 1")
+    masked_t = t.masked_fill(~mask, 0.0)
+
+    return masked_t.sum(dim=dim) / denom.clamp(min=1e-5)

ip_adapter/utils.py

@@ -0,0 +1,142 @@
+import torch
+import torch.nn.functional as F
+import numpy as np
+from PIL import Image
+
+BLOCKS = {
+    'content': ['down_blocks'],
+    'style': ["up_blocks"],
+
+}
+
+controlnet_BLOCKS = {
+    'content': [],
+    'style': ["down_blocks"],
+}
+
+
+def resize_width_height(width, height, min_short_side=512, max_long_side=1024):
+
+    if width < height:
+
+        if width < min_short_side:
+            scale_factor = min_short_side / width
+            new_width = min_short_side
+            new_height = int(height * scale_factor)
+        else:
+            new_width, new_height = width, height
+    else:
+
+        if height < min_short_side:
+            scale_factor = min_short_side / height
+            new_width = int(width * scale_factor)
+            new_height = min_short_side
+        else:
+            new_width, new_height = width, height
+
+    if max(new_width, new_height) > max_long_side:
+        scale_factor = max_long_side / max(new_width, new_height)
+        new_width = int(new_width * scale_factor)
+        new_height = int(new_height * scale_factor)
+    return new_width, new_height
+
+def resize_content(content_image):
+    max_long_side = 1024
+    min_short_side = 1024
+
+    new_width, new_height = resize_width_height(content_image.size[0], content_image.size[1],
+                                                min_short_side=min_short_side, max_long_side=max_long_side)
+    height = new_height // 16 * 16
+    width = new_width // 16 * 16
+    content_image = content_image.resize((width, height))
+
+    return width,height,content_image
+
+attn_maps = {}
+def hook_fn(name):
+    def forward_hook(module, input, output):
+        if hasattr(module.processor, "attn_map"):
+            attn_maps[name] = module.processor.attn_map
+            del module.processor.attn_map
+
+    return forward_hook
+
+def register_cross_attention_hook(unet):
+    for name, module in unet.named_modules():
+        if name.split('.')[-1].startswith('attn2'):
+            module.register_forward_hook(hook_fn(name))
+
+    return unet
+
+def upscale(attn_map, target_size):
+    attn_map = torch.mean(attn_map, dim=0)
+    attn_map = attn_map.permute(1,0)
+    temp_size = None
+
+    for i in range(0,5):
+        scale = 2 ** i
+        if ( target_size[0] // scale ) * ( target_size[1] // scale) == attn_map.shape[1]*64:
+            temp_size = (target_size[0]//(scale*8), target_size[1]//(scale*8))
+            break
+
+    assert temp_size is not None, "temp_size cannot is None"
+
+    attn_map = attn_map.view(attn_map.shape[0], *temp_size)
+
+    attn_map = F.interpolate(
+        attn_map.unsqueeze(0).to(dtype=torch.float32),
+        size=target_size,
+        mode='bilinear',
+        align_corners=False
+    )[0]
+
+    attn_map = torch.softmax(attn_map, dim=0)
+    return attn_map
+def get_net_attn_map(image_size, batch_size=2, instance_or_negative=False, detach=True):
+
+    idx = 0 if instance_or_negative else 1
+    net_attn_maps = []
+
+    for name, attn_map in attn_maps.items():
+        attn_map = attn_map.cpu() if detach else attn_map
+        attn_map = torch.chunk(attn_map, batch_size)[idx].squeeze()
+        attn_map = upscale(attn_map, image_size) 
+        net_attn_maps.append(attn_map) 
+
+    net_attn_maps = torch.mean(torch.stack(net_attn_maps,dim=0),dim=0)
+
+    return net_attn_maps
+
+def attnmaps2images(net_attn_maps):
+
+    #total_attn_scores = 0
+    images = []
+
+    for attn_map in net_attn_maps:
+        attn_map = attn_map.cpu().numpy()
+        #total_attn_scores += attn_map.mean().item()
+
+        normalized_attn_map = (attn_map - np.min(attn_map)) / (np.max(attn_map) - np.min(attn_map)) * 255
+        normalized_attn_map = normalized_attn_map.astype(np.uint8)
+        #print("norm: ", normalized_attn_map.shape)
+        image = Image.fromarray(normalized_attn_map)
+
+        #image = fix_save_attn_map(attn_map)
+        images.append(image)
+
+    #print(total_attn_scores)
+    return images
+def is_torch2_available():
+    return hasattr(F, "scaled_dot_product_attention")
+
+def get_generator(seed, device):
+
+    if seed is not None:
+        if isinstance(seed, list):
+            generator = [torch.Generator(device).manual_seed(seed_item) for seed_item in seed]
+        else:
+            generator = torch.Generator(device).manual_seed(seed)
+    else:
+        generator = None
+
+    return generator

requirements.txt

@@ -0,0 +1,18 @@
+diffusers==0.25.1
+torch==2.0.1
+torchaudio==2.0.2
+torchvision==0.15.2
+transformers==4.40.2
+accelerate
+safetensors
+einops
+spaces==0.19.4
+omegaconf
+peft
+huggingface-hub==0.24.5
+opencv-python
+insightface
+gradio
+controlnet_aux
+gdown
+peft