Repolode
投稿する
sczhou

sczhou/CodeFormer

Python tool for blind face restoration, enhancing blurry, low-quality or damaged faces in photos and videos using a codebook lookup transformer.

GitHub で見る
スター
18,185
フォーク
3,723
未解決の Issue
283
最終コミット
11 か月前

これは何？

CodeFormer restores faces in images and videos when the degradation type is unknown, a setting called blind face restoration. It is the official implementation of a NeurIPS 2022 paper that uses a learned codebook of facial priors and a transformer to predict clean facial details from degraded input.

The tool exposes a fidelity weight that trades off output quality against similarity to the original face, so users can favor cleaner results or closer identity. Beyond restoration, it includes scripts for colorizing black-and-white or faded face photos and for inpainting masked regions of aligned faces.

For whole images and video, it can enhance background regions with Real-ESRGAN and upsample the restored face. Pretrained models are downloaded via a provided script, and the project includes training code and configuration files.

主な機能

  • 01Blind face restoration for images and cropped, aligned 512x512 faces
  • 02Adjustable fidelity weight to balance quality and similarity to the input
  • 03Video enhancement from .mp4, .mov and .avi inputs
  • 04Face colorization for black-and-white or faded photos
  • 05Face inpainting for masked regions
  • 06Optional Real-ESRGAN background enhancement and face upsampling

ユースケース

  • Restoring old or low-resolution family photos
  • Enhancing blurry frames in older video clips
  • Fixing distorted faces in AI-generated art
  • Researchers comparing blind face restoration methods on aligned face datasets

トピック

#face-restoration#computer-vision#pytorch#super-resolution#image-enhancement#deep-learning#codebook#codeformer#face-enhancement#restoration

よくある質問

CodeFormer とは？
Python tool for blind face restoration, enhancing blurry, low-quality or damaged faces in photos and videos using a codebook lookup transformer. CodeFormer restores faces in images and videos when the degradation type is unknown, a setting called blind face restoration. It is the official implementation of a NeurIPS 2022 paper that uses a learned codebook of facial priors and a transformer to predict clean facial details from degraded input.
CodeFormer は無料で使えますか？ライセンスは？
CodeFormer は Other ライセンスのオープンソースです。商用製品で使う前にライセンス条件を確認してください。
CodeFormer は今もメンテナンスされていますか？
半年以上コミットなし。CodeFormer の最終コミットは 2025年11月18日 です。
CodeFormer は何に使えますか？
Restoring old or low-resolution family photos Enhancing blurry frames in older video clips Fixing distorted faces in AI-generated art Researchers comparing blind face restoration methods on aligned face datasets
CodeFormer はどこで入手できますか？
ソースコードとインストール手順は GitHub にあります：https://github.com/sczhou/CodeFormer

概要はプロジェクトの README をもとに AI が生成しています。GitHub の数値は毎日更新されます。

関連プロジェクト