NVlabs/Sana

Implements linear-complexity diffusion transformers with block causal attention to enable efficient training and inference across image and video generation tasks, supporting variants like SANA-Sprint (one-step) and SANA-Video (temporal synthesis). Integrates with major frameworks including HuggingFace Diffusers, ComfyUI, SGLang serving, and Cosmos-RL for reinforcement learning post-training. Provides complete training pipelines with multi-scale WebDataset support, 4-bit quantization, and ControlNet compatibility for production deployments.

5,000 stars. Actively maintained with 4 commits in the last 30 days.