OpenSubject

# OpenSubject Dataset OpenSubject is a video-derived large-scale corpus with 2.5M samples and 4.35M images for subject-driven generation and manipulation, as presented in the paper [OpenSubject: Leveraging Video-Derived Identity and Diversity Priors for Subject-driven Image Generation and Manipulation](https://huggingface.co/papers/2512.08294). ## Project Page & Code See the main repository for more details and code: [OpenSubject](https://github.com/LAW1223/OpenSubject) ## Dataset Structure ``` OpenSubject/ ├── Images_packages/ # Compressed image packages (tar.gz) │ ├── generation_input_images_*.tar.gz │ ├── generation_output_images_*.tar.gz │ ├── manipulation_input_images_*.tar.gz │ └── manipulation_output_images_*.tar.gz └── Jsonls/ # Annotation files ├── generation_merged.jsonl └── manipulation_merged.jsonl ``` ## Extracting Images After downloading, extract the image packages: ```bash python scripts/unzip_images/extract_images.py \ --packages_dir ./Images_packages \ --output_dir ./Images \ --num_workers 32 ``` This will create the following structure: ``` Images/ ├── generation/ │ ├── input_images/ │ └── output_images/ └── manipulation/ ├── input_images/ └── output_images/ ``` ## Sample Usage The CLI tool (`scripts/inference_cli.py` in the main repository) allows you to generate images directly from the command line. ### Basic Usage (Text-to-Image) Generate an image from a text prompt: ```bash python scripts/inference_cli.py \ --model_path /path/to/omnigen2_model \ --transformer_path /path/to/opensubject_model \ --prompt "a beautiful landscape with mountains and lakes" \ --output_path output.png \ --num_inference_step 50 \ --height 1024 \ --width 1024 ``` ### With Input Images (Image-to-Image) Generate an image with reference input images: ```bash python scripts/inference_cli.py \ --model_path /path/to/omnigen2_model \ --transformer_path /path/to/opensubject_model \ --prompt "transform the scene to sunset" \ --input_images input1.jpg input2.jpg \ --output_path result.png \ --num_inference_step 50 ``` ## License This dataset is released under the Apache 2.0 License.# OpenSubject 数据集 OpenSubject是一个源自视频的大规模语料库，包含250万条样本与435万张图像，专用于主体驱动的生成与操控任务，相关研究成果已发表于论文《OpenSubject：利用源自视频的身份与多样性先验知识实现主体驱动的图像生成与操控》，论文链接：https://huggingface.co/papers/2512.08294。 ## 项目页面与代码 如需获取更多细节及代码，请访问主仓库：[OpenSubject](https://github.com/LAW1223/OpenSubject) ## 数据集结构 OpenSubject/ ├── Images_packages/ # 压缩图像包（tar.gz格式） │ ├── generation_input_images_*.tar.gz │ ├── generation_output_images_*.tar.gz │ ├── manipulation_input_images_*.tar.gz │ └── manipulation_output_images_*.tar.gz └── Jsonls/ # 标注文件 ├── generation_merged.jsonl └── manipulation_merged.jsonl ## 图像解压方法 下载完成后，可通过以下命令解压图像包： bash python scripts/unzip_images/extract_images.py --packages_dir ./Images_packages --output_dir ./Images --num_workers 32 执行该命令后将生成如下目录结构： Images/ ├── generation/ │ ├── input_images/ │ └── output_images/ └── manipulation/ ├── input_images/ └── output_images/ ## 示例用法 该项目的命令行工具（CLI，位于主仓库的`scripts/inference_cli.py`）支持直接通过命令行生成图像。 ### 基础用法（文本到图像生成） 基于文本提示生成图像： bash python scripts/inference_cli.py --model_path /path/to/omnigen2_model --transformer_path /path/to/opensubject_model --prompt "a beautiful landscape with mountains and lakes" --output_path output.png --num_inference_step 50 --height 1024 --width 1024 ### 结合输入图像用法（图像到图像生成） 基于参考输入图像生成图像： bash python scripts/inference_cli.py --model_path /path/to/omnigen2_model --transformer_path /path/to/opensubject_model --prompt "transform the scene to sunset" --input_images input1.jpg input2.jpg --output_path result.png --num_inference_step 50 ## 许可证 本数据集采用Apache 2.0许可证发布。

搜集汇总

背景与挑战

背景概述

OpenSubject是一个视频衍生的大规模语料库，包含250万样本和435万张图像，专为基于主题的图像生成和操作而设计。数据集提供压缩的图像包和JSONL注释文件，并支持通过命令行工具进行图像生成和操作，遵循Apache 2.0许可证。