Sana
Linear diffusion transformer models for fast image and video generation
What it is
NVIDIA Labs' codebase and open weights for the SANA family of diffusion models, covering text-to-image, text-to-video, few-step distilled variants, a camera-controllable world model, and streaming video editing. It ships full training and inference pipelines, plus reinforcement-learning post-training recipes and an inference acceleration engine. Researchers and developers use it to generate high-resolution images and 720p video on modest GPUs, or to train and fine-tune their own variants.