sczhou/codeformer

CodeFormer is an AI-powered tool for restoring and enhancing faces in photos and videos. It was developed as a research project at Nanyang Technological University and presented at NeurIPS 2022, a top academic AI conference. The problem it solves: old or low-quality photos often have blurry, damaged, or degraded faces that are hard to view clearly. CodeFormer automatically sharpens and restores these faces using a technique called a "codebook lookup transformer", this means the AI has learned a library of high-quality face features and uses it to intelligently fill in missing or damaged details, rather than just blindly sharpening everything. Beyond basic face restoration, it also supports face colorization (turning black-and-white photos into color), face inpainting (filling in areas you mask out, like removing blemishes), and video enhancement. A key parameter called the "fidelity weight" lets you control a tradeoff between quality and faithfulness to the original, lower values give crisper, more enhanced results; higher values stay closer to what was originally there. You can try it through online demos on Hugging Face and Replicate without installing anything. For local use, it runs with PyTorch (a machine learning framework) and requires a CUDA-capable GPU (a type of graphics card used for AI processing). Input can be whole images or pre-cropped face photos, and the results are saved to an output folder. You'd use CodeFormer to restore family photos, fix AI-generated art with degraded faces, or enhance old video footage. It is written in Python.