Sana is NVIDIA’s open-source family of efficient generative models for high-resolution image synthesis, built around a linear diffusion transformer. Released by NVIDIA Research (NVlabs) under the Apache-2.0 license, the project has collected more than 9,400 stars on GitHub and was actively developed through 2026.
Most diffusion models get dramatically slower as resolution grows. Sana attacks that cost from both sides: a deep-compression autoencoder squeezes images into a much smaller latent space, and a linear-attention transformer replaces quadratic attention in the backbone, keeping quality while cutting compute. The result is a text-to-image model that can synthesize images up to 4096×4096, aligns strongly with text prompts, and still runs on consumer-grade GPUs — the Sana family has since grown toward faster distilled variants and video generation as well. Despite its small footprint (the original model is around 600M parameters), independent comparisons place its output quality close to models many times its size, which is exactly why it attracts teams that serve images at scale or deploy generation on edge hardware.
Sana is free and open source under the Apache-2.0 license, with model weights published for download; you supply your own compute.
The repository collects the Sana line in one place, including distilled few-step variants and the compact models most users start with, together with training, fine-tuning and inference scripts. Pre-trained weights are published for download, and the codebase has become a frequent reference point for research on efficient diffusion transformers.
It is for researchers and product teams who need high-resolution image generation that is cheap enough to run at scale or on modest hardware, and for anyone studying efficient diffusion architectures.