📄 Paper
🌐 Project Website

Code and project page for our paper "VODiff: Controlling Object Visibility Order in Text-to-Image Generation", accepted to CVPR 2025.

VODiff is a training-free framework that introduces object visibility order as a new controllable dimension in layout-to-image text-to-image (T2I) generation.

Compared with previous methods, which cannot explicitly control object occlusion, VODiff enables accurate generation of complex scenes with user-defined spatial and occlusion relationships via two core designs:

• Sequential Denoising Process (SDP): Synthesizes objects in layers, bottom to top, according to visibility order.
• Visibility-Order-Aware (VOA) Loss: Optimizes cross-attention maps to enforce correct spatial and occlusion constraints.

# Create a new conda environment
conda create --name vodiff python=3.9
conda activate vodiff

# Install dependencies
conda install -r requirements.txt

Download the pretrained model (e.g., GLIGEN) and place it in the checkpoints directory.

To run inference, use the provided Jupyter notebook, please modify these parts to define your own inputs.

caption = 'A car and a bike in front of a house.'
names_list = ['house', 'car', 'bike']  # Ordered by visibility (back to front)
layout = [(66, 197, 452, 390), (326, 358, 402, 432), (111, 347, 216, 431)]  # Corresponding bounding boxes

The names_list should be ordered by visibility, i.e., from background to foreground.

This project is built upon the following resources:

• Attention Refocusing: Our codebase is based on the foundational work provided by Attention Refocusing.

If you have any questions or issues, please feel free to open an issue or contact us.

This project is released under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International license.

🔗 License Details

If you find VODiff useful in your research, please consider citing us:

@inproceedings{liang2025vodiff,
  title={VODiff: Controlling Object Visibility Order in Text-to-Image Generation},
  author={Liang, Dong and Jia, Jinyuan and Liu, Yuhao and Ke, Zhanghan and Fu, Hongbo and Lau, Rynson WH},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={18379--18389},
  year={2025}
}