Sana is an open-source training and inference framework for high-resolution image and video generation. It uses Linear Attention in DiT to reduce computational cost and latency, enabling deployment on laptop GPUs with less than 8GB VRAM via 4-bit quantization.
Project overview
The framework addresses high computational cost and latency in diffusion models by using Linear Attention in DiT, allowing high-resolution generation to run on laptop GPUs with under 8GB VRAM.
Project type
Image & Vision · Video
Use cases
Image Processing · Video Creation
Deployment
Refer to project documentation
License
Apache-2.0
Best for
Developers, AI engineers, and researchers who need an open-source framework for model training and customization with efficient memory management.
Users operating on hardware with limited VRAM who require 4-bit quantization for high-resolution image generation.
Key capabilities
Achieves one or few-step generation via sCM distillation, reaching 0.1 seconds per 1024px image on an H100 GPU.
Provides real-time streaming video-to-video editing for 720p, minute-scale videos.
Includes NVFP4 Rollout and BF16 Training reinforcement learning post-training, achieving 4.64x faster convergence.
Limitations and risks
Full performance without quantization may require significant GPU resources, such as an H100, to achieve baseline latency targets.
The system requires pretrained SANA models loaded from HuggingFace or local storage to operate.
Setup and operation require coding knowledge to run environment scripts and utilize the library or CLI.
Getting started
Clone the repository from GitHub, run the provided environment setup shell script to install Python dependencies, and use SanaPipeline.from_pretrained to generate an image.