Sana

Overview

Sana is a free text-to-image generator from NVIDIA's research team that’s fast, efficient, and easy to use. It creates high-resolution images up to 4096×4096 pixels using text prompts and works on devices with a 16GB GPU.

With Sana, you can produce a 1024×1024 image in less than a second using a laptop GPU. The model’s standout features include a deep compression autoencoder, linear attention mechanisms, and a decoder-only text encoder. These innovations ensure high-quality results without slowing down performance.

Here’s how it works:

• Text-to-Image Creation: Turn text into detailed, high-resolution visuals.
• High-Resolution Output: Generate images up to 4096×4096 pixels.
• Speed and Efficiency: Create 1024×1024 images in under a second on a 16GB GPU.
• Clear Text in Images: Generates crisp text, including styles like neon signs and banners.
• Logo Design: Produces logos comparable to specialized AI tools.

License Scope
The training code got an Apache 2.0 update, but the model itself is still under NSCL v2-custom, which follows NVIDIA’s strict rules.

Key Restrictions

• Non-Commercial Use Only: You can only use it for research or evaluation with NVIDIA GPUs. NVIDIA keeps all commercial rights.
• Mandatory NSFW Filtering: Inputs and outputs must block anything explicit, harmful, or offensive.
• Liability for Violations: If your filtering fails and there’s a legal problem, you’re on the hook.

Termination Clause
If you violate the license, you lose all your rights to use the model.

Model has been censored quite a bit since its first release and now won't generate remotely NSFW outputs for prompts like 'sexy woman'.

NVIDIA made Sana to be lightweight yet quite powerful. Despite its smaller size, it competes with larger diffusion models and delivers results faster, making it great for creators, developers, and enthusiasts.

ComfyUI nodes for Sana are now available.

SANA-Sprint - the newer model - is all about speed without losing quality.

• 10× faster than top models like FLUX-Schnell
• 7.59 FID & 0.74 GenEval score—beats rivals in single-step quality
• Top-tier speed vs. quality balance—outshines SD3.5-Turbo and SDXL-DMD2

HyperNoise Sana Sprint

HyperNoise Sana Sprint 0.6B is a LoRA adapter made to improve how text-to-image models work. It tweaks the starting noise to help make better images that match what people want.

It runs on the Noise Hypernetworks setup. The trick is in a small hypernetwork that changes the noise going into a frozen model. Once it’s trained, it only needs one forward pass to make an image. No backprop or extra steps needed when generating.

They tested it on smaller models like SD-Turbo, SANA-Sprint and FLUX-Schnell. It got better GenEval scores, jumping from 0.70 to 0.75. Image quality held steady even across 32 steps. Results matched what you'd get with fancy prompt tricks but used way less power.

The model uses a KL-regularized noise goal and LoRA for a lightweight setup. It was trained in just one step but still works well for longer generations.

You can use it under the MIT license. The code, model, and docs are up on GitHub and Hugging Face. More info is on the Noise Hypernetworks site.