Repolode
Submit
NVlabs

NVlabs/Sana

Open-source codebase for efficient high-resolution image and video generation using linear diffusion transformers, with training and inference pipelines.

Open-source alternative to:MidjourneyRunway

View on GitHubWebsite
Stars
9,407
Forks
753
Open issues
138
Last commit
6 hours ago

What is it?

SANA is a diffusion-model codebase from NVIDIA Labs for generating high-resolution images and videos. Its core design replaces standard attention in the diffusion transformer with linear attention, along with a high-compression autoencoder (DC-AE), to reduce the cost of producing large outputs such as 4K images.

The repository covers a family of models: SANA for text-to-image up to 4K, SANA-1.5 for training-time and inference-time scaling, SANA-Sprint for one- or few-step generation, and several video models including SANA-Video, SANA-Video 2.0, a controllable world model (SANA-WM), and a streaming video-to-video editing model (SANA-Streaming). It also includes Sol-RL, a reinforcement-learning post-training recipe. Pretrained weights are published on Hugging Face and the models can be loaded through diffusers.

The project provides ComfyUI nodes, SGLang serving support, ControlNet, LoRA and DreamBooth fine-tuning, and 4-bit and 8-bit quantization for running on GPUs with limited memory.

Key features

  • 01Linear attention diffusion transformer for high-resolution image generation up to 4K.
  • 02Text-to-video and text-image-to-video models, including 720p outputs.
  • 03One- and few-step generation through sCM distillation.
  • 04ControlNet, LoRA and DreamBooth fine-tuning support.
  • 054-bit and 8-bit quantization for running on GPUs with limited VRAM.
  • 06Integrations with diffusers, ComfyUI and SGLang.

Use cases

  • Generating high-resolution images on a single consumer GPU.
  • Researching efficient diffusion transformer architectures.
  • Fine-tuning a text-to-image or video model on custom data with LoRA.
  • Serving a diffusion model behind an OpenAI-compatible API with SGLang.

Quick start

  1. 1

    Open the official documentation for installation and setup

Things to know

  • Generation needs a capable NVIDIA GPU; the README's demos reference 3090-class and H100 hardware.

Topics

#diffusion#text-to-image#text-to-video#linear-attention#transformers#pytorch#image-generation#video-generation#dit#linear-transformer

FAQ

What is Sana?
Open-source codebase for efficient high-resolution image and video generation using linear diffusion transformers, with training and inference pipelines. SANA is a diffusion-model codebase from NVIDIA Labs for generating high-resolution images and videos. Its core design replaces standard attention in the diffusion transformer with linear attention, along with a high-compression autoencoder (DC-AE), to reduce the cost of producing large outputs such as 4K images.
Is Sana free to use? What license does it have?
Sana is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is Sana still maintained?
Actively maintained. The last commit to Sana was on Oct 20, 2026.
What can I use Sana for?
Generating high-resolution images on a single consumer GPU. Researching efficient diffusion transformer architectures. Fine-tuning a text-to-image or video model on custom data with LoRA. Serving a diffusion model behind an OpenAI-compatible API with SGLang.
How do I install Sana?
Open the official documentation for installation and setup
Where can I get Sana?
The source code and installation instructions are on GitHub: https://github.com/NVlabs/Sana

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects