NVlabs/Sana

Project overview

SANA is an open-source training and inference framework for high-resolution image and video generation that uses Linear Attention in Diffusion Transformers to reduce computational costs. It is deployable on laptop GPUs with less than 8GB VRAM using 4-bit quantization, though full performance requires significant hardware such as an H100.

Project type

Model Development · Model Runtime · Image & Vision · Video

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 flexible hardware requirements.

Key capabilities

• Achieves one or few-step generation via sCM distillation, producing a 1024px image in 0.1 seconds on an H100 GPU.
• Provides real-time streaming video-to-video editing for 720p, minute-scale videos.
• Provides NVFP4 Rollout and BF16 Training Reinforcement Learning, achieving 4.64x faster convergence.

Limitations and risks

• Full performance without quantization may require significant GPU resources, such as an H100 for 0.1 second generation latency.

Getting started

• Clone the repository, navigate into the directory, run the environment_setup.sh shell script to install Python dependencies, and use SanaPipeline.from_pretrained to generate an image. Coding is required for setup and usage.

Evidence and sources

• README: **SANA** is an efficiency-oriented codebase for high-resolution image and video generation, providing complete training and inference pipelines.
• README: Deployable on laptop GPUs with **< 8GB VRAM** via 4-bit quantization.
• README: We introduce **SANA**, a series of efficient diffusion models for high-resolution image and video generation:
• README: git clone https://github.com/NVlabs/Sana.git cd Sana && ./environment_setup.sh sana
• README: - [✅] Training code - [✅] Inference code - [✅] Model zoo

Reviewed at: 2026-07-18T04:55:56Z