# SANA

An efficiency-oriented codebase for high-resolution image and video generation with a linear diffusion transformer, covering the SANA, SANA-Video, SANA-Sprint, SANA-WM and Sol-RL family of models with full training and inference pipelines.

SANA is an efficiency-oriented codebase for high-resolution image and video generation, providing complete training and inference pipelines. It covers a family of related models: SANA and SANA-1.5 for image generation, SANA-Sprint, SANA-Video and SANA-Video 2.0 for video generation, SANA-WM as a controllable world model, SANA-Streaming for real-time streaming video editing, and Sol-RL for reinforcement-learning training recipes.

SANA-Video 2.0 is an 8-second, 720p video model supporting both text-to-video and text-image-to-video generation with hybrid linear and softmax attention; a 4-step preview variant generates 720p video in four denoising steps. SANA-WM is a controllable world model supporting 720p, one-minute video generation with six-degrees-of-freedom camera control, and SANA-Streaming supports 720p, one-minute streaming video editing. The Sol Engine provides accelerated inference for models including MiniMax-H3, reporting several-times speedups on GB200, DGX Spark and GeForce RTX 5090 hardware without distillation, LoRA or a calibration pass.

It is written in Python, built on PyTorch, and released under the Apache-2.0 license.

Website: https://nvlabs.github.io/Sana/docs/