Z-Image-Turbo: the most downloaded image generator in the catalog, and where to run it
Apache 2.0, weights free, commercial use allowed, nothing to accept. At 6 billion parameters it fits within 16GB of video memory, so self-hosting needs a consumer graphics card and the Hugging Face diffusers library, and its authors report sub-second generation on an H800. In your browser it runs on cnaps studio, included in the basic plan.
It is the most downloaded image generator in this catalog, at 629,504 downloads in the last month against 591,129 for FLUX Schnell. What people come for is a photorealistic image in eight model passes, with English or Chinese text rendered inside the picture when the prompt asks for it.
Two costs, both stated by its own authors. Output diversity is rated low, so one prompt tends to return variations on one picture, and no fine-tuning path is offered. It also publishes no score of its own. Seven image generation models sit in the cnaps studio catalog, and this is the one to reach for when you want a photograph from a sentence.
What it actually is
A single-stream diffusion transformer. Many generators run text and image through separate branches that meet later. This one concatenates the text tokens, a semantic reading of the image and the compressed image itself into one sequence that passes through one stack, which its authors argue spends more of the parameter budget on the work itself.
Speed comes from a distillation method called Decoupled DMD, which cuts generation to eight function evaluations. In cnaps studio you set diffusion steps from 5 to 30, a seed, and output height and width, which default to 1024 by 1024. The default of 9 steps is the setting that produces those eight passes.
Input is a text prompt in English or Chinese. There is no reference image, no mask and no editing mode.
Try Z-Image-Turbo in cnaps studio →
The published numbers
The model card publishes no figure of its own. It points to a human preference evaluation on Alibaba's AI Arena, where people pick the better of two images, and calls the result highly competitive and state of the art among open models. The leaderboard appears only as a picture, with no number in the text.
The figures that exist come from a rival. GLM-Image's model card benchmarks it in the same tables as GLM-Image itself. On CVTG-2K, a text-rendering test, Z-Image-Turbo scores 0.8585 word accuracy against 0.9116. On DPG-Bench, which checks whether long, dense prompts are followed, it scores 84.86 against 84.78.
Both pairs come from one table measured one way, so they are comparable, though the team doing the measuring built the other model. Read together: level with GLM-Image on following prompts, behind it when a picture needs a line of legible text.
How it compares to the other image generators here
Apart from the GLM-Image pair above, these four share no common test, so the table describes what each model is and leaves quality to your own prompts. All four licenses allow commercial use; SANA-Sprint 1.6B's adds a use policy that a product must pass on to its end users.
Reach for Z-Image-Turbo for a photorealistic image from a prompt, on the model most people already use. Reach for SANA-Sprint 1.6B when speed decides, FLUX Schnell for a larger text-only model, and GLM-Image when the picture needs readable words or an existing image needs changing. For blending several references, OmniGen2 or FLUX.2 KLEIN 4B are the better fit.
What to chain it with
Text in, one picture out, which puts it at the start of any chain.
Two chains cover what it lacks. For changes, pass the result to QWEN-2511 for an instructed edit or to QWEN-Inpaint to replace one masked region, keeping a composition you like. For size, PiSA-SR or SMFANet+ enlarges the 1024-pixel output for print, and BiRefNet (Subject) cuts the subject out for a new background.
Open Z-Image-Turbo in cnaps studio, run one prompt four times at the default nine steps, and see how far the results drift before you plan a batch around it.
FAQ
Is Z-Image-Turbo free for commercial use?
Yes. The weights are released under Apache 2.0, which allows commercial use, modification and redistribution without a separate agreement. Self-hosting needs a graphics card with 16GB of video memory and the diffusers library. In cnaps studio it is included in the basic plan, so there is nothing to install or configure.
Why do my Z-Image-Turbo images look alike?
Its authors rate its output diversity as low, so running one prompt with different seeds tends to return variations on a single composition. To get genuinely different results, change the prompt itself: the subject's position, the lighting, the camera angle or the setting. Changing the seed alone moves the details and leaves the idea in place.
Can Z-Image-Turbo put text inside an image?
Yes, in English or Chinese. On the CVTG-2K text-rendering benchmark published on GLM-Image's model card, it scores 0.8585 word accuracy against 0.9116 for GLM-Image. For a poster or a label where every word has to be legible, GLM-Image is the stronger choice in the cnaps studio catalog.
