SANA-Sprint 1.6B: a 1024-pixel image in one to four steps, and where to run it
Start with the license. SANA-Sprint 1.6B is distributed under the NVIDIA Open Model License, which allows commercial use. One condition comes with it: the bundled text encoder carries Google's prohibited-use policy, and a service built on the model has to pass that policy through to its end users. Self-hosting is light: 1.6 billion parameters, and the authors report 0.31 seconds per 1024-pixel image on an RTX 4090. In your browser it runs on cnaps studio, included in the basic plan.
It is built for speed. Diffusion models usually refine a picture over twenty or more passes, and this one was distilled to produce a 1024-pixel image in one to four steps. Its own card also measures it head to head against FLUX Schnell, another model in this catalog, and it comes out ahead on both published scores.
The authors list the costs plainly. It does not reach perfect photorealism, it cannot render complex legible text, and fingers can come out wrong. Seven image generation models sit in the cnaps studio catalog, and this is the one with the lowest default step count.
What it actually is
A linear diffusion transformer, the efficient-attention design of the SANA family, distilled with continuous-time consistency distillation. That training lets one set of weights work anywhere from one to four steps, so there is no separate checkpoint for each setting.
Two supporting parts carry much of the load. A frozen 2-billion-parameter language model reads the prompt, and an autoencoder called DC-AE compresses the image 32 times spatially before the transformer works on it, which is how a 1.6-billion-parameter model handles 1024 pixels.
In cnaps studio you set diffusion steps from 1 to 4, defaulting to 2, a seed, and width and height, defaulting to 1024. Under advanced options a guidance scale, defaulting to 4.5, sets how strictly the prompt is followed. Prompts can be English or Chinese.
Try SANA-Sprint 1.6B in cnaps studio →
The published numbers
Two quality scores, both at one step, both set against FLUX Schnell on the same card. FID, which measures how close generated images sit to real photographs, is 7.59 against 7.94, where lower is better. GenEval, which checks whether the objects, counts and colours a prompt asks for actually appear, is 0.74 against 0.71, where higher is better.
Speed is the bigger claim: 0.1 seconds per 1024-pixel image against 1.1 seconds for FLUX Schnell on the same data-centre GPU, roughly ten times faster.
All of it comes from NVIDIA's own card, and the quality margins are small. The fair reading is level or slightly better at about a tenth of the time, measured by the side that benefits from the result.
How it compares to the other image generators here
Only the first two columns share a test. All four licenses allow commercial use, though they differ in detail: the other three are Apache 2.0, and SANA-Sprint 1.6B's license adds the text encoder's use policy for end users.
Reach for SANA-Sprint 1.6B for fast drafts, prototypes and high-volume runs where speed decides. Reach for FLUX Schnell or Z-Image-Turbo for a larger text-only model, FLUX.2 KLEIN 4B when you need to edit or blend references, and GLM-Image when the picture needs legible words.
What to chain it with
Text in, one picture out, which puts it at the start of a chain.
Its speed suits a draft-then-refine pattern: iterate on the prompt here, then take the version you like to Z-Image-Turbo or FLUX Schnell for the final image. To fix one detail in a draft, QWEN-Inpaint replaces a masked region, and PiSA-SR or SMFANet+ enlarges the result for print.
Open SANA-Sprint 1.6B in cnaps studio, run one prompt at one step and again at four, and see what the extra three passes buy.
FAQ
Can I use SANA-Sprint 1.6B commercially?
Yes. It is distributed under the NVIDIA Open Model License, which allows commercial use. The bundled text encoder carries Google's prohibited-use policy, so a product built on it has to pass that policy through to its end users in its terms. In cnaps studio it runs in the browser, included in the basic plan.
How fast is SANA-Sprint 1.6B?
Its authors report 0.1 seconds per 1024-pixel image on an H100 and 0.31 seconds on an RTX 4090, against 1.1 seconds for FLUX Schnell on the same H100. In cnaps studio the default is two steps, and you can set anywhere from one to four.
Can SANA-Sprint 1.6B put text inside an image?
Not reliably. Its own model card lists complex legible text among its limitations, alongside imperfect photorealism and badly drawn fingers. For signs, posters or labels where every word has to read correctly, GLM-Image is the model in the cnaps studio catalog built for that job.
