OSDFace: One-Step Diffusion Model for Face Restoration

Jingkai Wang, Jue Gong, Lin Zhang, Zheng Chen, Xing Liu, Hong Gu, Yutong Liu, Yulun Zhang, and Xiaokang Yang, "One-Step Diffusion Model for Face Restoration", CVPR, 2025

🔥🔥🔥 News

2025-12-23: Inference code and pretrained models are released.
2025-04-23: Results are released. (Synthetic dataset: CelebA-Test; Real-world datasets: Wider-Test, LFW-Test, and WebPhoto-Test)
2025-02-27: Congratulations! OSDFace is accepted to CVPR 2025.
2024-11-25: This repo is released.

Abstract: Diffusion models have demonstrated impressive performance in face restoration. Yet, their multi-step inference process remains computationally intensive, limiting their applicability in real-world scenarios. Moreover, existing methods often struggle to generate face images that are harmonious, realistic, and consistent with the subject’s identity. In this work, we propose OSDFace, a novel one-step diffusion model for face restoration. Specifically, we propose a visual representation embedder (VRE) to better capture prior information and understand the input face. In VRE, low-quality faces are processed by a visual tokenizer and subsequently embedded with a vector-quantized dictionary to generate visual prompts. Additionally, we incorporate a facial identity loss derived from face recognition to further ensure identity consistency. We further employ a generative adversarial network (GAN) as a guidance model to encourage distribution alignment between the restored face and the ground truth. Experimental results demonstrate that OSDFace surpasses current state-of-the-art (SOTA) methods in both visual quality and quantitative metrics, generating high-fidelity, natural face images with high identity consistency.

⚒️ TODO

[x] Release inference code and pretrained models
[ ] Release training code

🔗 Contents

[x] Environment
[x] Datasets
[x] Models
[x] Inference
[ ] Training
[x] Results
[x] Citation
[ ] Acknowledgements

<a name="environment"></a>⚒️ Environment

Create conda environment with Python 3.10. (conda create -n osdface python=3.10)
Install PyTorch 2.4.0. (pip install torch==2.4.0 torchvision==0.19.0 torchaudio==2.4.0 --index-url https://mirror.sjtu.edu.cn/pytorch-wheels/cu121)
Install requirements.txt. (pip install -r requirements.txt)

<a name="datasets"></a>📊 Datasets

Training Dataset:

We use FFHQ and FFHQR for training. All of them are resized to 512x512.

Testing Dataset:

We use CelebA-Test as the synthetic dataset. The download link can refer to DAEFR.

We use Wider-Test, LFW-Test, and WebPhoto-Test as the real-world datasets. The download links can refer to DAEFR and VQFR.

Thanks a lot for their contribution to Face Restoration.

<a name="inference"></a>🧮 Inference

Download the pretrained model from Google Drive. Put models in pretrained/, and put the test data in data/
Run the inference bash test.sh.

💬 Tips

Since the offcial Stable Diffusion 2.1 is not avaliable, you could visit ModelScope to download it and use it locally.

If you meet an error from diffusers:

File "... lib/python3.10/site-packages/diffusers/utils/dynamic_modules_utils.py", line 28, in <module>
    from huggingface_hub import cached_download, hf_hub_download, model_info

you can edit the dynamic_modules_utils.py like:

Replace line 24 from from huggingface_hub import cached_download, hf_hub_download, model_info to from huggingface_hub import hf_hub_download, model_info.
Replace line 287 from resolved_module_file = cached_download( to resolved_module_file = hf_hub_download(.

<a name="results"></a>🔎 Results

We achieved state-of-the-art performance on synthetic and real-world datasets. Detailed results can be found in the paper, and the full results could be downloaded here: Google Drive or OneDrive.

<details> <summary>&ensp;Quantitative Comparisons (click to expand) </summary> <li> Results in Table 1 on synthetic dataset (CelebA-Test) from the main paper. <img src="images/quan-celeba.png" > </li> <li> Results in Table 2 on real-world datasets (Wider-Test, LFW-Test, WebPhoto-Test) from the main paper. <img src="images/quan-realworld.png" > </li> </details> <details open> <summary>&ensp;Visual Comparisons (click to expand) </summary> <li> Results in Figure 5 on synthetic dataset (CelebA-Test) from the main paper. <img src="images/vis-main-celeba.png" > </li> <li> Results in Figure 6 on real-world dataset (Wider-Test, LFW-Test, WebPhoto-Test) from the main paper. <img src="images/vis-main-realworld.png" > </li> </details> <details> <summary style="margin-left: 2rem;">&ensp;More Comparisons on Synthetic Dataset... </summary> <li style="margin-left: 2rem;"> Results in Figure 4, 5, 6 on synthetic dataset (CelebA-Test) from supplemental material. <img src="images/vis-supp-celeba-1.png" > <img src="images/vis-supp-celeba-2.png" > <img src="images/vis-supp-celeba-3.png" > </li> </details> <details> <summary style="margin-left: 2rem;">&ensp;More Comparisons on Real-World Dataset... </summary> <li style="margin-left: 2rem;"> Results in Figure 7, 8, 9, 10 on real-world datasets (Wider-Test, LFW-Test, WebPhoto-Test) from supplemental material. <img src="images/vis-supp-wider.png" > <img src="images/vis-supp-lfw.png" > <img src="images/vis-supp-webphoto.png" > <img src="images/vis-supp-realworld.png" > </li> </details>

<a name="citation"></a>📎 Citation

If you find the code helpful in your research or work, please cite the following paper(s).

@InProceedings{wang2025osdface,
    author    = {Wang, Jingkai and Gong, Jue and Zhang, Lin and Chen, Zheng and Liu, Xing and Gu, Hong and Liu, Yutong and Zhang, Yulun and Yang, Xiaokang},
    title     = {{OSDFace}: One-Step Diffusion Model for Face Restoration},
    booktitle = {Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR)},
    month     = {June},
    year      = {2025},
    pages     = {12626-12636}
}

<a name="acknowledgements"></a>💡 Acknowledgements

[TBD]

OSDFace

Install / Use

README