NVIDIA's kitchen-sink framework for speeding up diffusion models

What it does

FastGen is NVIDIA’s training framework for turning slow, many-step diffusion models into fast few-step generators. It wraps consistency models, distribution-matching distillation, knowledge distillation, and several newer variants into one configurable system. The target models span images (EDM, SD, SDXL, Flux, Qwen-Image) and video (WAN, CogVideoX, Cosmos), with tasks including text-to-image, image-to-video, and video-to-video.

The interesting bit

Rather than championing one distillation technique, FastGen treats them as interchangeable plugins. The same trainer loop runs CM, sCM, DMD2, Self-Forcing, MeanFlow, and others against the same network backends. That design makes it a testbed for comparing methods head-to-head, though the README notes not every method-network pairing is implemented yet.

Key highlights

• Supports models at 10B+ parameters with FSDP2 sharding
• Hydra-style config overrides without editing files (- key=value)
• Built-in EMA, checkpointing, and W&B logging via callbacks
• Docker environment provided; otherwise conda + pip install
• Planned release of distilled CIFAR-10 and ImageNet checkpoints

Caveats

• No pretrained student checkpoints available yet; you train your own
• README warns that “not all combinations of methods and networks are currently supported”
• Video model support appears newer and less battle-tested than image pipelines

Verdict

Researchers and engineers who need to benchmark distillation methods or productionize fast diffusion/video generation should grab this. Casual users looking for drop-in faster models should wait for the promised checkpoint release or look elsewhere.