Hybrid Layout Control for Diffusion Transformer: Fewer Annotations, Superior Aesthetics

Keming Wu^1*, Junwen Chen^3*, Zhanhao Liang^2*, Yinuo Wang^1*, Ji Li⁵, Chao Zhang⁴, Bin Wang¹, Yuhui Yuan^6*

¹Tsinghua University   ²The Australian National University   ³The University of Electro-Communications Tokyo   ⁴Peking University   ⁵Microsoft  ⁶Canva

^*Work done at Microsoft Research Asia

🌟ICCV 2025🌟

🔥 News

• [2025/7/20] Repository is initialized.
• [2025/6/26] 🎉🎉🎉 HybridLayout is accepted by ICCV 2025! 🎉🎉🎉

🚧 TODO List

• Release inference code and pretrained model
• Release training code

Table of Contents

• Environment Setup
• Testing

Environment Setup

1. Create Conda Environment

conda create -n hybrid_layout python=3.10 -y
conda activate hybrid_layout

2. Install Dependencies

git clone https://github.com/KemingWu/HybridLayout.git
cd HybridLayout
pip install uv
uv pip install --pre -U xformers
uv pip install diffusers==0.31.0 transformers==4.44.0
uv pip install mmengine

3. Login to Hugging Face

huggingface-cli login

✨ Quick Start

Use our inference.ipynb to simply have a try

🔥 Contact

If you have any questions, please feel free to contact Keming Wu and Yuhui Yuan.

📬 Citation

If you find this code useful in your research, please consider citing:

@inproceedings{wu2025hybrid,
  title={Hybrid Layout Control for Diffusion Transformer: Fewer Annotations, Superior Aesthetics},
  author={Wu, Keming and Chen, Junwen and Liang, Zhanhao and Wang, Yinuo and Li, Ji and Zhang, Chao and Wang, Bin and Yuan, Yuhui},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={17930--17940},
  year={2025}
}