Generative Refocusing: Flexible Defocus Control from a Single Image

This model, Generative Refocusing, presented in the paper Generative Refocusing: Flexible Defocus Control from a Single Image, offers a novel two-step process for depth-of-field control from a single image. It uses DeblurNet to recover all-in-focus images from various inputs and BokehNet for creating controllable bokeh. The method leverages semi-supervised training, combining synthetic paired data with unpaired real bokeh images, and achieves state-of-the-art performance in defocus deblurring, bokeh synthesis, and refocusing benchmarks, allowing text-guided adjustments and custom aperture shapes.

⚡ Quick Start

Follow the steps below to set up the environment and run the inference demo.

1. Installation

Clone the repository:

git clone git@github.com:rayray9999/Genfocus.git
cd Genfocus

Environment setup:

conda create -n Genfocus python=3.12
conda activate Genfocus

Install requirements:

pip install -r requirements.txt

2. Download Weights

You can download the pre-trained models using the following commands. Ensure you are in the Genfocus root directory.

# 1. Download main models to the root directory
wget https://huggingface.co/nycu-cplab/Genfocus-Model/resolve/main/bokehNet.safetensors
wget https://huggingface.co/nycu-cplab/Genfocus-Model/resolve/main/deblurNet.safetensors

# 2. Setup checkpoints directory and download auxiliary model
mkdir -p checkpoints
cd checkpoints
wget https://huggingface.co/nycu-cplab/Genfocus-Model/resolve/main/checkpoints/depth_pro.pt
cd ..

3. Run Gradio Demo

Launch the interactive web interface locally:

Note: The project uses FLUX.1-dev. You must request access and authenticate locally before running the demo.

python demo.py

The demo will be accessible at http://127.0.0.1:7860 in your browser.

Citation

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

@article{Genfocus2025,
  title={Generative Refocusing: Flexible Defocus Control from a Single Image},
  author={Tuan Mu, Chun-Wei and Huang, Jia-Bin and Liu, Yu-Lun},
  journal={arXiv preprint arXiv:2512.16923},
  year={2025}
}