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

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

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 ..

Launch the interactive web interface locally:

Note: This project uses FLUX.1-dev. You must request access and authenticate locally before running the demo.
⚠️ VRAM warning: GPU memory usage can be high. On an NVIDIA A6000, peak usage may reach ~45GB depending on resolution and settings.
Tip (Advanced Settings): If you want faster inference, try lowering num_inference_steps, resizing the input in Advanced Settings.

python demo.py

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

You can also run inference directly via the command line for our different models.

Restore sharp details from blurry images.

⚠️ Update Notice: If you downloaded the weights before our recent inference command updates, please re-download the new deblurNet.safetensors to ensure best performance.

python Inference_deblurNet.py \
  --input inference_example/Blurry_example.jpg \
  --output Deblurred_output.png 

• --input / -i: Path to the input blurry image.
• --output / -o: Path to save the output image.

Add realistic bokeh effects to an All-In-Focus (AIF) image using a depth/focus mask.

python Inference_bokehNet.py \
  --input inference_example/AIF_example.png \
  --mask inference_example/AIF_mask.png \
  --depth inference_example/AIF_example_pred.npy \
  --k_value 15 \
  --output Bokeh_output.png 

• --input / -i: Path to the All-In-Focus input image.
• --mask / -m: Path to the in-focus mask image.
• --point / -p: Focus point x,y on the ORIGINAL image (e.g., 512,300).
• --depth / -d: Path to a pre-computed depth map (.npy file). If not provided, Depth Pro is automatically used.
• --k_value / -k: Blur strength K.
• --output / -o: Path to save the output image.

A variant of DeblurNet that utilizes a pre-deblurring module for heavily degraded images.

Note: Please download the specific weight for this variant before running the inference:

wget https://huggingface.co/nycu-cplab/Genfocus-Model/resolve/main/deblurNet_with_pre_deblur.safetensors

python Inference_deblurNet_with_pre_deblur.py \
  --input inference_example/Blurry_example.jpg \
  --pre_deblur_input inference_example/Blurry_example_pre_deblur.jpg \
  --output Deblurred_output_with_pre_deblur.png 

• --input / -i: Path to the input blurry image.
• --pre_deblur_input: Path to the pre-processed/pre-deblurred image.
• --output / -o: Path to save the output image.

A ComfyUI implementation of Genfocus is available, thanks to Eric Rollei!
Check it out here: 👉 comfyui-refocus

We are actively working on improving this project. Current progress:

• Upload Model Weights
• Release HF Demo & Gradio Code (with tiling tricks for high-res images)
• Release Inference Code (Support for adjustable parameters/settings)
• Release Benchmark data
• Release Training Code and Data

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}
}

For any questions or suggestions, please open an issue or contact me at raytm9999.cs09@nycu.edu.tw.