Self-Supervised Selective-Guided Diffusion Model for Old-Photo Face Restoration

Wenjie Li¹  Xiangyi Wang¹  Heng Guo¹  Guangwei Gao²  Zhanyu Ma¹

¹Beijing University of Posts and Telecommunications  ²Nanjing University of Science and Technology 

🚩 Accepted to NeurIPS 2025

• [Project]   [arXiv]   [Appendix]•

⭐ If SSDiff is helpful to your paper or project, please consider star this repo or cite our paper. Thanks! 🤗

---

📮 Update

• 2025.02.11: Datas are relased.
• 2025.12.07: Codes are relased.
• 2025.12.03: Checkpoints and scripts are relased.
• 2025.12.02: Config files are relased.
• 2025.10.11: This repo is created.

🔧 TODO

• Relase restoration results.
• Relase our constructed old-photo face dataset, VintageFace.
• Relase checkpoint and script for old photo face restoration.
• Relase codes and config files.
• Release paper on arxiv.

Dependencies and Installation

• Pytorch >= 1.7.1
• CUDA >= 10.1
• Other required packages in requirements.txt

# git clone this repository
git clone https://github.com/PRIS-CV/SSDiff
cd SSDiff

# create new anaconda env
conda create -n SSDiff python=3.8 -y
conda activate SSDiff

# install python dependencies
conda install mpi4py
pip3 install -r requirements.txt
pip install -e .

Quick Inference

Download Pre-trained Models:

Download the pretrained face generation diffusion model iddpm_ffhq512_ema500000.pth from [Google Drive] to models folder, the pretrained face parsing model resnet34.pt from [Google Drive] to models/face_parsing folder, the pretrained style transfer model photo_image.pt from [Google Drive] to CAP_VSTNet/checkpoints folder.

Download the facelib and dlib pretrained models from [Releases | OneDrive] to the weights/facelib folder and weights folder, separately. You can manually download the pretrained models OR download by running the following command:

python scripts/download_pretrained_models.py facelib
python scripts/download_pretrained_models.py dlib (only for dlib face detector)

Download the pretrained restore model codeformer.pth and pretrained VQGAN model vqgan_code1024.pth from [Releases | OneDrive] to the models/restorer folder and models/vqgan folder, separately. You can manually download the pretrained models OR download by running the following command:

python scripts/download_pretrained_models.py CodeFormer

Download VintageFace Test dataset:

Download the real-world old face photo test set that collected from the Internet called VintageFace from [Google Drive], including old face photos with their corresponding processed masks.

Prepare Testing Data:

You can put the testing images in the inputs/TestWhole folder. If you would like to test on cropped and aligned faces, you can put them in the inputs/cropped_faces folder. You can get the cropped and aligned faces by running the following command:

# you may need to install dlib via: conda install -c conda-forge dlib
python extract_faces.py -i [input folder] -o [output folder]

• If scratches exist, a folder for mask(s) mask_dir must be specified with the name of each mask image corresponding to that of each input image. Each input mask shoud be a binary map with white pixels representing masked regions. To obtain a scratch map automatically, we recommend using the scratch detection model. One may also further adjust the scratch map with an image editing app (e.g., Photoshop).

Visual results:

You can download visual results from Google Drive.

Testing:

🧑🏻 Old-Photo Face Restoration (cropped and aligned face)

# For generating reference pseudo-labelled faces (512x512)
python inference_fake_label.py --task old_photo_restoration_pseudo --in_dir [input_image_path]  --mask_dir [mask_path] --out_dir [output_image_path] --guidance_scale 0.001 --seed 4321

# For breakage, fading, and blur faces (512x512)
# --self_dir refers to the path of the pseudo-labelled faces generated in the previous stage.
python inference_final.py --task old_photo_restoration --in_dir [input_image_path]  --mask_dir [mask_path] --out_dir [output_image_path] --guidance_scale 0.0035 --self_dir [pseudo_labelled_faces_path] --seed 4321

🧑🏻 Face Restoration (cropped and aligned face)

# For only blur faces (512x512), You may choose not to use pseudo-labelled faces.
python inference_final.py --task restoration --in_dir [input_image_path] --out_dir [output_image_path] --guidance_scale 0.05

Citation

If our work is useful for your research, please consider citing:

@inproceedings{li2025self,
    title={Self-Supervised Selective-Guided Diffusion Model for Old-Photo Face Restoration},
    author={Li, Wenjie and Wang, Xiangyi and Guo, Heng and Gao, Guangwei and Ma, Zhanyu},
    booktitle={NeurIPs},
    year={2025}
}

Contact

If you have any questions, please feel free to reach me out at lewj2408@gmail.com.