🚀

Image Generation - SANA-Sprint 1.6B

🛠

Ultra-Fast One-Step Text-to-Image — 1024px Images in 1 to 4 Steps

Ultra-fast one-step text-to-image generation (SANA-Sprint 1.6B). Generates 1024px images in just 1 – 4 diffusion steps, versus 20+ steps for a typical diffusion model. On an H100, produces a 1024×1024 image in ~0.1 seconds — roughly 10× faster than FLUX-schnell while scoring better on image-quality benchmarks. Runs at ~0.31 s on an RTX 4090.

What it does

SANA-Sprint is a small, extremely fast text-to-image model from NVIDIA and MIT Han Lab. Give it a text prompt — a scene description, a subject, a mood — and it generates a 1024×1024 image in a fraction of a second. The key trick behind its speed: instead of the usual 20+ "denoising steps" that a normal diffusion model needs, SANA-Sprint has been trained to produce a good image in as few as 1 step (and never more than 4). On benchmark tests it beats FLUX-schnell — a well-known "fast" image model — on both image quality (FID 7.59 vs 7.94) and prompt-following (GenEval 0.74 vs 0.71), while running roughly ten times faster. It supports both English and Chinese prompts and produces images at 1024×1024 by default (multi-scale height and width also supported). Important: the underlying model is released under NVIDIA's Open Model License and is intended for research use — see the License section below before deploying it commercially..

Problem it solves

  • Interactive, real-time generation – At ~0.1s per 1024×1024 image on an H100 (and ~0.31s on an RTX 4090), the model is fast enough for interactive prompt iteration, live demos, and AIPC-style consumer apps
  • Ultra-low compute cost – 1-4 diffusion steps instead of 20+ means dramatically less GPU time per image — good for batch generation, low-cost inference, or hardware-constrained deployment
  • Beats the fast-tier competition – Outperforms FLUX-schnell on both image quality (FID) and prompt following (GenEval) while running ~10× faster
  • Modern quality at small size – 1.6B parameters (BF16) is small enough to run comfortably on consumer GPUs while still producing high-quality 1024px output
  • Bilingual prompts – Native English + Chinese prompt support out of the box
  • Real-time ControlNet – The SANA-Sprint family also includes a ControlNet variant for structure-guided generation at real-time speeds (0.25s on H100 for 1024px)

Input/Output

  • Input: Text prompt (English or Chinese)
    • image
  • Options:
    • Diffusion steps (1 – 4, default 2)
      • Fewer steps = faster generation; higher steps trade a little speed for better quality
      • The model was specifically distilled to work well anywhere in this 1-4 step range
    • Random seed (-1 = random, default 1)
      • Set a fixed integer to reproduce the exact same image
    • Output width (px) — default 1024
    • Output height (px) — default 1024
    • Advanced Options:
      • Classifier-free guidance (CFG) scale — default 4.5
        • Higher values push the model to follow the prompt more strictly; lower values allow more creative freedom
  • Output: Generated image
    • 1024 × 1024 native resolution (multi-scale H × W supported)
    • Photorealistic or stylized depending on the prompt
      • image

Accuracy & Speed

Speed on H100
~0.1 seconds per 1024×1024 image — 10× faster than FLUX-schnell
Speed on RTX 4090
~0.31 seconds per 1024×1024 image — consumer-GPU friendly
Diffusion steps
1 – 4 (typical: 2) — vs. 20+ for standard diffusion models
FID score (image quality, lower = better)
7.59 — beats FLUX-schnell (7.94) at 1 step
GenEval score (prompt-following, higher = better)
0.74 — beats FLUX-schnell (0.71) at 1 step
Model size
1.6B parameters (BF16)
Native resolution
1024px base with multi-scale H × W
Languages
English and Chinese
ControlNet variant
Real-time structure-guided generation available in the SANA-Sprint family (~0.25 s on H100

Technical Details

Architecture
Linear Diffusion Transformer (from the SANA family) with one-step / few-step consistency distillation
Distillation Method
Continuous-time Consistency Distillation (sCM) — a training-free approach that avoids costly retraining while enabling 1–4 step generation
Step-Adaptive Design
Unified single model that produces high-quality output across the 1 – 4 step range (no need for separate 1-step / 4-step checkpoints)
Text Encoder
Gemma 2 2B Instruct (google/gemma-2-2b-it) — frozen, pre-trained
Latent Feature Encoder
DC-AE (Deep Compression Autoencoder) — 32× spatial compression
Parameters
1.6B
Precision
bfloat16 (BF16)
Base Resolution
1024px (multi-scale height × width supported)
Supported Frameworks
NVLabs SANA repository (github.com/NVlabs/Sana); Hugging Face Diffusers integration available
Family Members
Original SANA (many-step), SANA-Sprint 0.6B and 1.6B (few-step), SANA-Sprint ControlNet
Paper
"SANA-Sprint: One-Step Diffusion with Continuous-Time Consistency Distillation" — arXiv:2503.09641 (Mar 12, 2025)

Compliance & Provenance

Provider
NVIDIA (open weights)
Provider type
Specialized
License
NVIDIA Open Model License
EU AI Act risk class
Limited Risk
Art. 50 transparency
Required — outputs are marked. See AI Policy §2.
Region availability
Available globally
Training data summary
Provider model card / training data

For more on how we classify models and mark outputs, see our AI Policy.

License Note (Important)

  • License: NVIDIA Open Model License Agreement
  • Intended use: Research purposes only per NVIDIA's model card
  • Deployment consideration: Before using SANA-Sprint 1.6B in a commercial product, review the NVIDIA Open Model License Agreement carefully — its terms differ from the Apache 2.0 / MIT licenses on most other models in the studio, and it may restrict certain commercial uses

Model Source