NVlabs/FastGen

Implements multiple model distillation and acceleration techniques—including consistency models, distribution matching, and knowledge distillation—unified under a configurable PyTorch framework. Supports cross-modal generation tasks (text-to-image, image-to-video, video-to-video) with architectures ranging from EDM and Stable Diffusion to Flux and CogVideoX, scaling to ≥10B parameters via DDP and FSDP2 distributed training. Uses Hydra configuration system for experiment management and integrates W&B logging with modular design for custom networks and datasets.

632 stars. Actively maintained with 3 commits in the last 30 days.