nvidia / sana-sprint-1.6b

SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation

  • Public
  • 6.6K runs
  • GitHub
  • Weights
  • Paper
  • License

logo

🐱 Sana Model Card

Demos

Demo Video of SANA-Sprint Demo Video of SANA-Sprint

Training Pipeline

teaser_page1

Model Efficiency

teaser_page1

SANA-Sprint is an ultra-efficient diffusion model for text-to-image (T2I) generation, reducing inference steps from 20 to 1-4 while achieving state-of-the-art performance. Key innovations include: (1) A training-free approach for continuous-time consistency distillation (sCM), eliminating costly retraining; (2) A unified step-adaptive model for high-quality generation in 1-4 steps; and (3) ControlNet integration for real-time interactive image generation. SANA-Sprint achieves 7.59 FID and 0.74 GenEval in just 1 step — outperforming FLUX-schnell (7.94 FID / 0.71 GenEval) while being 10× faster (0.1s vs 1.1s on H100). With latencies of 0.1s (T2I) and 0.25s (ControlNet) for 1024×1024 images on H100, and 0.31s (T2I) on an RTX 4090, SANA-Sprint is ideal for AI-powered consumer applications (AIPC).

Source code is available at https://github.com/NVlabs/Sana.

Model Description

Model Sources

For research purposes, we recommend our generative-models Github repository (https://github.com/NVlabs/Sana), which is more suitable for both training and inference MIT Han-Lab provides free SANA-Sprint inference. - Repository: https://github.com/NVlabs/Sana - Demo: https://nv-sana.mit.edu/sprint - Guidance: https://github.com/NVlabs/Sana/asset/docs/sana_sprint.md

Uses

Direct Use

The model is intended for research purposes only. Possible research areas and tasks include

  • Generation of artworks and use in design and other artistic processes.
  • Applications in educational or creative tools.
  • Research on generative models.
  • Safe deployment of models which have the potential to generate harmful content.

  • Probing and understanding the limitations and biases of generative models.

Excluded uses are described below.

Out-of-Scope Use

The model was not trained to be factual or true representations of people or events, and therefore using the model to generate such content is out-of-scope for the abilities of this model.

Limitations and Bias

Limitations

  • The model does not achieve perfect photorealism
  • The model cannot render complex legible text
  • fingers, .etc in general may not be generated properly.
  • The autoencoding part of the model is lossy.

Bias

While the capabilities of image generation models are impressive, they can also reinforce or exacerbate social biases.