GOBUNU/NTIRE2025_RealWorld_Face_Restoration

NTIRE 2025 Real-World Face Restoration

This repository provides the pretrained weights used by the public implementation of the NTIRE 2025 Challenge on Real-World Face Restoration at CVPR 2025.

The challenge focuses on restoring high-quality facial details from real-world degraded images while maintaining identity consistency. The released files cover the participating methods that can be run through the challenge repository, together with shared pretrained components required by those methods.

Resources

• —Challenge website
• —Project page
• —Code and inference instructions
• —CVPR 2025 workshop paper
• —arXiv report

Repository structure

model_zoo/   # Team-specific checkpoints and model assets
pretrained/  # Shared pretrained components used by the evaluation pipelines

Some participating teams chose not to publish their models, so this repository contains only the checkpoints made available for release.

Download

Install the Hugging Face Hub client:

pip install -U huggingface_hub

Download the weights into the root of the cloned challenge repository:

hf download GOBUNU/NTIRE2025_RealWorld_Face_Restoration \
  --include "model_zoo/**" "pretrained/**" \
  --local-dir .

The resulting directory layout should contain ./model_zoo and ./pretrained. See the challenge repository for environment setup, model selection, inference commands, and evaluation instructions.

Intended use

These checkpoints are intended for research, reproducibility, and evaluation of real-world face-restoration methods submitted to the NTIRE 2025 challenge. Performance and resource requirements vary across participating solutions.

Face restoration can alter fine facial details. Outputs should not be treated as ground-truth evidence, and users should consider privacy, consent, and the potential impact of generated or reconstructed facial content.

Security note

Some PyTorch checkpoints may use pickle-based serialization. Download and load model files only in a trusted environment, review the corresponding implementation, and avoid executing untrusted code.

License and acknowledgements

The challenge code repository is released under the MIT License. Individual methods, pretrained components, and checkpoints may be subject to their respective upstream licenses or release terms. Users are responsible for reviewing the applicable terms before redistribution or commercial use.

This release includes work contributed by the challenge participants and incorporates components from the broader face-restoration and image-quality-assessment research communities. Please consult the code repository and the cited paper for detailed acknowledgements.

Citation

@inproceedings{ntiface25face,
  title     = {NTIRE 2025 Challenge on Real-World Face Restoration: Methods and Results},
  author    = {Chen, Zheng and Wang, Jingkai and Liu, Kai and Gong, Jue and Sun, Lei and Wu, Zongwei and Timofte, Radu and Zhang, Yulun and others},
  booktitle = {CVPR Workshops},
  year      = {2025}
}