anycost-gan

Why we included this project

Anycost GAN is the reference implementation of a CVPR 2021 paper that tackles a real annoyance in GAN-based image editing: the full generator takes about three seconds per image, which is too slow for interactive work. The generator can instead run at reduced channel and resolution settings, producing a preview that stays visually close to the final result while you adjust sliders, then rendering the high-quality output once you finalize. That makes it a good fit for anyone building face or car editing tools who wants responsive feedback without giving up final fidelity. The repo ships pre-trained generators, encoders, and editing directions for FFHQ and LSUN Car, plus a Colab notebook and demo video, so you can judge the approach quickly. It is a research codebase rather than a maintained product, so plan to adapt the PyTorch code to your own pipeline.