Unified Style and Subject-Driven Generation via Disentangled and Reward Learning

Shaojin Wu, Mengqi Huang, Yufeng Cheng, Wenxu Wu, Jiahe Tian, Yiming Luo, Fei Ding, Qian He
UXO Team
Intelligent Creation Lab, Bytedance

🔥 News

• 2025.08.28 🔥 The demo of USO is released. Try it Now! ⚡️
• 2025.08.28 🔥 Update fp8 mode as a primary low vmemory usage support (please scroll down). Gift for consumer-grade GPU users. The peak Vmemory usage is ~16GB now.
• 2025.08.27 🔥 The inference code and model of USO are released.
• 2025.08.27 🔥 The project page of USO is created.
• 2025.08.27 🔥 The technical report of USO is released.

📖 Introduction

Existing literature typically treats style-driven and subject-driven generation as two disjoint tasks: the former prioritizes stylistic similarity, whereas the latter insists on subject consistency, resulting in an apparent antagonism. We argue that both objectives can be unified under a single framework because they ultimately concern the disentanglement and re-composition of “content” and “style”, a long-standing theme in style-driven research. To this end, we present USO, a Unified framework for Style driven and subject-driven GeneratiOn. First, we construct a large-scale triplet dataset consisting of content images, style images, and their corresponding stylized content images. Second, we introduce a disentangled learning scheme that simultaneously aligns style features and disentangles content from style through two complementary objectives, style-alignment training and content–style disentanglement training. Third, we incorporate a style reward-learning paradigm to further enhance the model’s performance.

⚡️ Quick Start

🔧 Requirements and Installation

Install the requirements

## create a virtual environment with python >= 3.10 <= 3.12, like
python -m venv uso_env
source uso_env/bin/activate
## or
conda create -n uso_env python=3.10 -y
conda activate uso_env

## install torch
## recommended version:
pip install torch==2.4.0 torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu124 

## then install the requirements by you need
pip install -r requirements.txt # legacy installation command

Then download checkpoints:

# 1. set up .env file
cp example.env .env

# 2. set your huggingface token in .env (open the file and change this value to your token)
HF_TOKEN=your_huggingface_token_here

#3. download the necessary weights (comment any weights you don't need)
pip install huggingface_hub
python ./weights/downloader.py

• IF YOU HAVE WEIGHTS, COMMENT OUT WHAT YOU DON'T NEED IN ./weights/downloader.py

✍️ Inference

• Start from the examples below to explore and spark your creativity. ✨

# the first image is a content reference, and the rest are style references.

# for subject-driven generation
python inference.py --prompt "The man in flower shops carefully match bouquets, conveying beautiful emotions and blessings with flowers. " --image_paths "assets/gradio_examples/identity1.jpg" --width 1024 --height 1024
# for style-driven generation
# please keep the first image path empty
python inference.py --prompt "A cat sleeping on a chair." --image_paths "" "assets/gradio_examples/style1.webp" --width 1024 --height 1024
# for style-subject driven generation (or set the prompt to empty for layout-preserved generation)
python inference.py --prompt "The woman gave an impassioned speech on the podium." --image_paths "assets/gradio_examples/identity2.webp" "assets/gradio_examples/style2.webp" --width 1024 --height 1024
# for multi-style generation
# please keep the first image path empty
python inference.py --prompt "A handsome man." --image_paths "" "assets/gradio_examples/style3.webp" "assets/gradio_examples/style4.webp" --width 1024 --height 1024

# for low vram:
python inference.py --prompt "your propmt" --image_paths "your_image.jpg" --width 1024 --height 1024 --offload --model_type flux-dev-fp8 

• You can also compare your results with the results in the assets/gradio_examples folder.

• For more examples, visit our project page or try the live demo.

🌟 Gradio Demo

python app.py

For low vmemory usage, please pass the --offload and --name flux-dev-fp8 args. The peak memory usage will be 16GB (Single reference) ~ 18GB (Multi references).

# please use FLUX_DEV_FP8 replace FLUX_DEV
export FLUX_DEV_FP8="YOUR_FLUX_DEV_PATH"

python app.py --offload --name flux-dev-fp8

🌈 More examples

We provide some prompts and results to help you better understand the model. You can check our paper or project page for more visualizations.

Subject/Identity-driven generation

If you want to place a subject into new scene, please use natural language like "A dog/man/woman is doing...". If you only want to transfer the style but keep the layout, please an use instructive prompt like "Transform the style into ... style". For portraits-preserved generation, USO excels at producing high skin-detail images. A practical guideline: use half-body close-ups for half-body prompts, and full-body images when the pose or framing changes significantly.

Style-driven generation

Just upload one or two style images, and use natural language to create want you want. USO will generate images follow your prompt and match the style you uploaded.

Style-subject driven generation

USO can stylize a single content reference with one or two style refs. For layout-preserved generation, just set the prompt to empty. Layout-preserved generation

Layout-shifted generation

📄 Disclaimer

We open-source this project for academic research. The vast majority of images used in this project are either generated or from open-source datasets. If you have any concerns, please contact us, and we will promptly remove any inappropriate content. Our project is released under the Apache 2.0 License. If you apply to other base models, please ensure that you comply with the original licensing terms.

This research aims to advance the field of generative AI. Users are free to create images using this tool, provided they comply with local laws and exercise responsible usage. The developers are not liable for any misuse of the tool by users.

🚀 Updates

For the purpose of fostering research and the open-source community, we plan to open-source the entire project, encompassing training, inference, weights, dataset etc. Thank you for your patience and support! 🌟

• Release technical report.
• Release github repo.
• Release inference code.
• Release model checkpoints.
• Release huggingface space demo.
• Release training code.
• Release dataset.

Citation

If USO is helpful, please help to ⭐ the repo.

If you find this project useful for your research, please consider citing our paper:

@article{wu2025uso,
    title={USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning},
    author={Shaojin Wu and Mengqi Huang and Yufeng Cheng and Wenxu Wu and Jiahe Tian and Yiming Luo and Fei Ding and Qian He},
    year={2025},
    eprint={2508.18966},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
}