FastGen v0.1.0

Release of NVIDIA FastGen, a PyTorch framework for building fast generative models using various distillation and acceleration techniques.

Highlights

Modular, model-agnostic design
Built with Hydra and OmegaConf, allowing methods, networks, and data configurations to be composed independently and extended with additional models and datasets.

Distillation and training methods
CM, sCM, TCM, MeanFlow, DMD2, f-Distill, LADD, CausVid, Self-Forcing, SFT, KD, and more.

Network support
EDM, EDM2, DiT, Stable Diffusion 1.5, SDXL, Flux, Qwen-Image, WAN, CogVideoX, and Cosmos Predict2.

Image and video workflows
Support for T2I, T2V, I2V, V2V, and autoregressive long-video generation.

Production-scale training
DDP and FSDP2, mixed precision, EMA, Weights & Biases integration, automatic resume, and WebDataset with local or S3-backed data.

Reproducibility

The FastGen results reported in the README files included in this release were obtained using this version.