🚀 CSGO：文本到图像生成中的内容 - 风格组合

本项目 CSGO 是论文 CSGO: Content-Style Composition in Text-to-Image Generation 的官方 PyTorch 实现。我们持续更新和改进此仓库，若您发现任何问题或有建议，欢迎提出问题或提交拉取请求（PR）💖。

🚀 快速开始

1. 克隆代码并准备环境

git clone https://github.com/instantX-research/CSGO
cd CSGO

# 使用 conda 创建环境
conda create -n CSGO python=3.9
conda activate CSGO

# 使用 pip 安装依赖
# 适用于 Linux 和 Windows 用户
pip install -r requirements.txt

2. 下载预训练权重（即将推出）

从 HuggingFace 下载预训练权重是最简单的方法：

# 首先，确保安装了 git-lfs，详见：https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage
git lfs install
# 克隆并移动权重
git clone https://huggingface.co/InstantX/CSGO

我们的方法与 SDXL、VAE、ControlNet 和 图像编码器 完全兼容。请将它们下载并放置在 ./base_models 文件夹中。

3. 推理

import torch
from ip_adapter.utils import resize_content
import numpy as np
from ip_adapter.utils import BLOCKS as BLOCKS
from ip_adapter.utils import controlnet_BLOCKS as controlnet_BLOCKS
from PIL import Image
from diffusers import (
    AutoencoderKL,
    ControlNetModel,
    StableDiffusionXLControlNetPipeline,

)
from ip_adapter import CSGO


device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

base_model_path =  "./base_models/stable-diffusion-xl-base-1.0"  
image_encoder_path = "./base_models/IP-Adapter/sdxl_models/image_encoder"
csgo_ckpt = "./CSGO/csgo.bin"
pretrained_vae_name_or_path ='./base_models/sdxl-vae-fp16-fix'
controlnet_path = "./base_models/TTPLanet_SDXL_Controlnet_Tile_Realistic"
weight_dtype = torch.float16


vae = AutoencoderKL.from_pretrained(pretrained_vae_name_or_path,torch_dtype=torch.float16)
controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16,use_safetensors=True)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    base_model_path,
    controlnet=controlnet,
    torch_dtype=torch.float16,
    add_watermarker=False,
    vae=vae
)
pipe.enable_vae_tiling()


target_content_blocks = BLOCKS['content']
target_style_blocks = BLOCKS['style']
controlnet_target_content_blocks = controlnet_BLOCKS['content']
controlnet_target_style_blocks = controlnet_BLOCKS['style']

csgo = CSGO(pipe, image_encoder_path, csgo_ckpt, device, num_content_tokens=4,num_style_tokens=32,
                          target_content_blocks=target_content_blocks, target_style_blocks=target_style_blocks,controlnet_adapter=True,
                              controlnet_target_content_blocks=controlnet_target_content_blocks, 
                              controlnet_target_style_blocks=controlnet_target_style_blocks,
                              content_model_resampler=True,
                              style_model_resampler=True,

                              )

style_name = 'img_1.png'
content_name = 'img_0.png'
style_image = Image.open("../assets/{}".format(style_name)).convert('RGB')
content_image = Image.open('../assets/{}'.format(content_name)).convert('RGB')

caption ='a small house with a sheep statue on top of it'

num_sample=4

# 图像驱动的风格迁移
images = csgo.generate(pil_content_image= content_image, pil_style_image=style_image,
                           prompt=caption,
                           negative_prompt= "text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                           content_scale=1.0,
                           style_scale=1.0,
                           guidance_scale=10,
                           num_images_per_prompt=num_sample,
                           num_samples=1,
                           num_inference_steps=50,
                           seed=42,
                           image=content_image.convert('RGB'),
                           controlnet_conditioning_scale=0.6,
                          )

# 文本编辑驱动的风格合成
caption='a small house'
images = csgo.generate(pil_content_image= content_image, pil_style_image=style_image,
                           prompt=caption,
                           negative_prompt= "text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                           content_scale=1.0,
                           style_scale=1.0,
                           guidance_scale=10,
                           num_images_per_prompt=num_sample,
                           num_samples=1,
                           num_inference_steps=50,
                           seed=42,
                           image=content_image.convert('RGB'),
                           controlnet_conditioning_scale=0.4,
                          )

# 文本驱动的风格合成
caption='a cat'
# 如果内容图像仍对生成结果有干扰，可将内容图像设置为空图像。
# content_image =Image.fromarray(np.zeros((content_image.size[0],content_image.size[1], 3), dtype=np.uint8)).convert('RGB')

images = csgo.generate(pil_content_image= content_image, pil_style_image=style_image,
                           prompt=caption,
                           negative_prompt= "text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                           content_scale=1.0,
                           style_scale=1.0,
                           guidance_scale=10,
                           num_images_per_prompt=num_sample,
                           num_samples=1,
                           num_inference_steps=50,
                           seed=42,
                           image=content_image.convert('RGB'),
                           controlnet_conditioning_scale=0.01,
                          )

✨ 主要特性

• 我们的 CSGO 实现了 图像驱动的风格迁移、文本驱动的风格化合成以及文本编辑驱动的风格化合成。
• 更多结果请访问我们的 主页 🔥

📦 安装指南

克隆代码并准备环境

git clone https://github.com/instantX-research/CSGO
cd CSGO

# 使用 conda 创建环境
conda create -n CSGO python=3.9
conda activate CSGO

# 使用 pip 安装依赖
# 适用于 Linux 和 Windows 用户
pip install -r requirements.txt

下载预训练权重（即将推出）

从 HuggingFace 下载预训练权重是最简单的方法：

# 首先，确保安装了 git-lfs，详见：https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage
git lfs install
# 克隆并移动权重
git clone https://huggingface.co/InstantX/CSGO

💻 使用示例

基础用法

# 克隆代码并准备环境
git clone https://github.com/instantX-research/CSGO
cd CSGO

# 使用 conda 创建环境
conda create -n CSGO python=3.9
conda activate CSGO

# 使用 pip 安装依赖
# 适用于 Linux 和 Windows 用户
pip install -r requirements.txt

# 下载预训练权重（即将推出）
# 首先，确保安装了 git-lfs，详见：https://docs.github.com/en/repositories/working-with-files/managing-large-files/installing-git-large-file-storage
git lfs install
# 克隆并移动权重
git clone https://huggingface.co/InstantX/CSGO

# 推理代码
import torch
from ip_adapter.utils import resize_content
import numpy as np
from ip_adapter.utils import BLOCKS as BLOCKS
from ip_adapter.utils import controlnet_BLOCKS as controlnet_BLOCKS
from PIL import Image
from diffusers import (
    AutoencoderKL,
    ControlNetModel,
    StableDiffusionXLControlNetPipeline,

)
from ip_adapter import CSGO


device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

base_model_path =  "./base_models/stable-diffusion-xl-base-1.0"  
image_encoder_path = "./base_models/IP-Adapter/sdxl_models/image_encoder"
csgo_ckpt = "./CSGO/csgo.bin"
pretrained_vae_name_or_path ='./base_models/sdxl-vae-fp16-fix'
controlnet_path = "./base_models/TTPLanet_SDXL_Controlnet_Tile_Realistic"
weight_dtype = torch.float16


vae = AutoencoderKL.from_pretrained(pretrained_vae_name_or_path,torch_dtype=torch.float16)
controlnet = ControlNetModel.from_pretrained(controlnet_path, torch_dtype=torch.float16,use_safetensors=True)
pipe = StableDiffusionXLControlNetPipeline.from_pretrained(
    base_model_path,
    controlnet=controlnet,
    torch_dtype=torch.float16,
    add_watermarker=False,
    vae=vae
)
pipe.enable_vae_tiling()


target_content_blocks = BLOCKS['content']
target_style_blocks = BLOCKS['style']
controlnet_target_content_blocks = controlnet_BLOCKS['content']
controlnet_target_style_blocks = controlnet_BLOCKS['style']

csgo = CSGO(pipe, image_encoder_path, csgo_ckpt, device, num_content_tokens=4,num_style_tokens=32,
                          target_content_blocks=target_content_blocks, target_style_blocks=target_style_blocks,controlnet_adapter=True,
                              controlnet_target_content_blocks=controlnet_target_content_blocks, 
                              controlnet_target_style_blocks=controlnet_target_style_blocks,
                              content_model_resampler=True,
                              style_model_resampler=True,

                              )

style_name = 'img_1.png'
content_name = 'img_0.png'
style_image = Image.open("../assets/{}".format(style_name)).convert('RGB')
content_image = Image.open('../assets/{}'.format(content_name)).convert('RGB')

caption ='a small house with a sheep statue on top of it'

num_sample=4

# 图像驱动的风格迁移
images = csgo.generate(pil_content_image= content_image, pil_style_image=style_image,
                           prompt=caption,
                           negative_prompt= "text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                           content_scale=1.0,
                           style_scale=1.0,
                           guidance_scale=10,
                           num_images_per_prompt=num_sample,
                           num_samples=1,
                           num_inference_steps=50,
                           seed=42,
                           image=content_image.convert('RGB'),
                           controlnet_conditioning_scale=0.6,
                          )

高级用法

# 文本编辑驱动的风格合成
caption='a small house'
images = csgo.generate(pil_content_image= content_image, pil_style_image=style_image,
                           prompt=caption,
                           negative_prompt= "text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                           content_scale=1.0,
                           style_scale=1.0,
                           guidance_scale=10,
                           num_images_per_prompt=num_sample,
                           num_samples=1,
                           num_inference_steps=50,
                           seed=42,
                           image=content_image.convert('RGB'),
                           controlnet_conditioning_scale=0.4,
                          )

# 文本驱动的风格合成
caption='a cat'
# 如果内容图像仍对生成结果有干扰，可将内容图像设置为空图像。
# content_image =Image.fromarray(np.zeros((content_image.size[0],content_image.size[1], 3), dtype=np.uint8)).convert('RGB')

images = csgo.generate(pil_content_image= content_image, pil_style_image=style_image,
                           prompt=caption,
                           negative_prompt= "text, watermark, lowres, low quality, worst quality, deformed, glitch, low contrast, noisy, saturation, blurry",
                           content_scale=1.0,
                           style_scale=1.0,
                           guidance_scale=10,
                           num_images_per_prompt=num_sample,
                           num_samples=1,
                           num_inference_steps=50,
                           seed=42,
                           image=content_image.convert('RGB'),
                           controlnet_conditioning_scale=0.01,
                          )

📚 详细文档

模型细节

我们目前发布了两个模型权重。

管道流程

功能展示

演示示例

内容 - 风格组合

循环翻译

文本驱动的风格合成

文本编辑驱动的风格合成

🔧 技术细节

本项目使用 PyTorch 实现，与 SDXL、VAE、ControlNet 和 图像编码器 完全兼容。

📄 许可证

本项目采用 Apache-2.0 许可证。

致谢

本项目由 InstantX 团队开发，保留所有版权。

引用 💖

如果您发现 CSGO 对您的研究有用，欢迎给此仓库加星 🌟 并使用以下 BibTeX 引用我们的工作：

@article{xing2024csgo,
       title={CSGO: Content-Style Composition in Text-to-Image Generation}, 
       author={Peng Xing and Haofan Wang and Yanpeng Sun and Qixun Wang and Xu Bai and Hao Ai and Renyuan Huang and Zechao Li},
       year={2024},
       journal = {arXiv 2408.16766},
}

仓库 Star 历史