mit-han-lab/anycost-gan

Employs sampling-based multi-resolution training, adaptive-channel pruning, and generator-conditioned discriminators to enable StyleGAN2-based generators to operate across diverse computational budgets while maintaining output consistency. Supports both uniform and flexible channel configurations at variable resolutions (e.g., 256-1024px) and channel ratios, enabling 5x faster interactive preview generation for real-time editing workflows. Built on PyTorch with pre-trained models for face (FFHQ) and car synthesis, includes encoder and semantic editing directions, and provides distributed evaluation tools using Horovod for FID/PPL metrics.

782 stars. No commits in the last 6 months.