NVlabs/Sana

SANA: Efficient High-Resolution Image Synthesis with Linear Diffusion Transformer

What it solves

SANA addresses the high computational cost and slow speeds typically associated with generating high-resolution images and videos. It provides an efficiency-oriented framework that allows users to create high-quality visual content—ranging from 1024px images to 4K resolution and long-form videos—significantly faster and with lower hardware requirements than traditional large-scale diffusion models.

How it works

SANA utilizes a collection of optimized diffusion models and architectures designed for speed. It employs techniques like distillation (such as sCM), quantization (4-bit and 8-bit support), and advanced inference engines to reduce memory usage and increase generation speed. The codebase supports a full lifecycle of tasks, including training, fine-tuning (LoRA), and high-performance inference through integrations with tools like ComfyUI, SGLang, and Hugging Face Diffusers.

Who it’s for

This project is for researchers, developers, and creators who need to generate high-resolution visual media but are constrained by hardware limitations or the need for real-time performance. It is particularly useful for those looking to deploy efficient image or video generation pipelines on consumer-grade hardware or optimized server environments.

Highlights

  • High-Resolution Output: Supports image generation up to 4K and long-form video generation.
  • Extreme Efficiency: Models are significantly smaller and faster than comparable industry standards, with some configurations generating 1024px images in as little as 0.1 seconds.
  • Versatile Ecosystem: Full support for training, inference, and fine-tuning, with native integrations for ComfyUI, SGLang, and Hugging Face.
  • Broad Capabilities: Includes specialized models for streaming video, controllable world modeling, and one-step refinement for high-quality video drafts.

Related

  • Dispatch
    SANA-WM: Breaking the Minute Barrier in Open-Source World Modeling

    NVIDIA researchers introduce SANA-WM, a 2.6B parameter model capable of generating high-fidelity, 720p, minute-long videos with precise camera control on a single GPU.

  • Project
    thu-ml/TurboDiffusion

    TurboDiffusion is a video generation acceleration framework that speeds up end-to-end diffusion generation by 100-200x using attention acceleration and timestep distillation.

  • Project
    nunchux-ai/nunchaku

    A high-performance inference engine that uses SVDQuant to enable 4-bit quantization for diffusion models, significantly reducing VRAM usage and increasing inference speed.

  • Project
    Tencent-Hunyuan/HunyuanVideo-1.5

    A lightweight 8.3B parameter video generation model that enables high-quality text-to-video and image-to-video synthesis on consumer-grade GPUs.

  • Dispatch
    Nano Banana 2 Lite Release: High-Speed, Cost-Efficient Image Generation

    Google DeepMind has released Nano Banana 2 Lite (Gemini 3.1 Flash-Lite Image), a distilled version of the Nano Banana 2 model designed for low-latency generation and reduced operational costs.