CodeFormer

Badly degraded faces are hard to restore because many plausible faces could explain the same damaged input. CodeFormer, a NeurIPS 2022 method from S-Lab at Nanyang Technological University, narrows that ambiguity by first learning a discrete codebook of high-quality facial parts, then using a transformer to predict which codebook entries best describe a degraded face. A fidelity weight between 0 and 1 trades crisp detail against faithfulness to the original. The repository provides PyTorch inference scripts for aligned 512 pixel face crops, whole photographs with optional Real-ESRGAN background upscaling, and video clips, plus separate checkpoints for colorizing faded monochrome faces and filling masked areas. Training code and configs are included. Setup needs a conda environment, the bundled BasicSR package, downloaded weights, and ideally a CUDA GPU. Its S-Lab license, version 1.0, permits non-commercial use only, and commercial deployment requires contacting the authors, who note that several CodeFormer-branded websites are unofficial. Good for anyone repairing old portraits or AI art.