Run Alibaba's undistilled Z-Image Base in your browser. One prompt returns up to 4 images that differ in framing, people, light and style, so you can pick a direction before you refine it. No GPU or ComfyUI setup.
Variety by design
Base keeps the variety that Turbo trades for polish, so every image in a run is a new take on your prompt. Both grids below are illustrations laid out as four-image runs.
Open prompt
Name only the subject and the moment when you don't know the shot yet. Base answers with different camera distances, different people, different light and sometimes a different medium; keep the take that reads best and write its framing and style into the next prompt.
Fixed style
Pin the backdrop, outfit, pose and light, and Base still changes who appears from image to image. Use it to audition faces for a character or a campaign before you settle on one.
Pick Turbo for polish and Base for variety: Turbo is distilled from Base and rated higher for visual quality, while Base is rated higher for variety. Ratings come from Tongyi's Z-Image model card and GitHub README (2026), which cover Base and Turbo only; credits are per image on Oeneo AI.
| Compare | Z-Image Base | Z-Image Turbo | Qwen Image 3 |
|---|---|---|---|
| Model | Undistilled 6B baseJanuary 27, 2026 | Distilled from BaseNovember 26, 2025 | Separate Qwen modelJuly 21, 2026 |
| Images per run | 1–4 | 1 | 1–4 |
| Credits per image | 2 | 1 | 8 |
| Reference images | Text only | Text only | Up to 3 |
| Variety across images | Higher | Lower | — |
| Visual quality rating | High | Very high | — |
| CFG & negative prompt | In the model; fixed here | Not used | — |
| Best for | Directions, style tests, casting | Fast photoreal shots, short text | Lettering, reference edits |
Write a short description, not a tag list. Every detail you name is one Base stops varying, so spell out what must stay and leave the rest open.
Subject, action and place in plain sentences
“An older woman lifts a honeycomb frame toward the camera on a green hillside” gives Base more to work with than a string of keywords.
Where the subject sits and how big it is
Position words such as “small in the left third” or “close-up from a low angle” move the layout, which helps when a banner or cover needs empty space for a title.
The medium, once you've picked a direction
Put it first (“two-color linocut print”, “gouache storybook illustration”) so the look stops drifting from image to image.
Where the light comes from
“Low sun behind the subject” or “bright noon light” sets the mood; leave it out when you want Base to try different light.
Quality tags
“Masterpiece, 8K, best quality” add noise, not detail. Describe the materials and the light you want instead.
Lists of what to avoid
“No text, no blur” inside the main prompt can pull those things into the image. Describe the clean result you want instead.
Complicated hand poses
Hands and limbs are a known weak spot. Keep poses simple, say what the hands hold, and pick the cleanest image from the run.
Example prompt: A red-haired cyclist in a yellow rain jacket rides along a seawall, small in the left third of a wide frame, open sky on the right for a headline, low morning sun from behind, gouache storybook illustration with visible brush texture.
Each model has its own page and runs in the same workbench on one credit balance.
AlibabaRerun the chosen direction as a cleaner photoreal image.
View model
AlibabaAdd on-image lettering or build the next version from reference images.
View model
New
GoogleConsistent characters, product placement, exact short text
View model
New
Black Forest LabsComposed scenes from a prompt or references
View modelRemove the background, upscale, vectorize or denoise any result, all on the same credit balance.
Transparent PNG or WebP with clean hair and product edges, ready for a listing or a new composition.
2× / 4× / 8× with recovered texture instead of smeared pixels, so a 1K draft can become a print.
Logos, icons and flat illustrations become editable SVG that scales to any size.
Remove grain and compression artifacts from low-light photos and scans while keeping real detail.
Z-Image Base is the 6B-parameter, undistilled text-to-image model that Alibaba's Tongyi-MAI lab released on January 27, 2026; Z-Image Turbo is distilled from it. Base gives up some polish in exchange for more variety between images and a wider range of styles.
Z-Image is the official name, and fal and the community call the same model Z-Image Base or ZIB. Z-Image-Omni-Base (generation plus editing) and Z-Image-Edit are separate models that Tongyi lists as “to be released”; neither was out as of October 2026, so this page runs the original Base.
Yes. The generator at the top of this page runs Z-Image Base in the cloud, so you don't need a BF16-capable graphics card, a model download or a ComfyUI workflow. Write a prompt, pick the ratio and number of images, and the credit cost shows before you generate.
You can try it free: the 10 credits you get at sign-up cover 5 Z-Image Base images. After that, each image costs 2 credits at any ratio.
Not on Oeneo AI yet. The model supports all three, and Tongyi suggests 28–50 steps with a CFG of 3–5 plus a negative prompt, but here you only set the prompt, ratio, number of images and format. Write what you want to see instead of what to avoid.
Not reliably. Base is built to vary faces and builds from image to image, and it doesn't accept reference images. For a recurring character, use a model that takes reference images, such as Qwen Image 3.
No. Base only turns text into images, and Tongyi's editing model, Z-Image-Edit, isn't released. To change an existing picture, use Qwen Image 3 with reference images, or the Oeneo AI tools for upscaling, background removal and denoising.
Not yet. Tongyi positions Base as the version for LoRA and fine-tuning, but Oeneo AI runs only the official base weights, without LoRA loading or training.
Yes. Tongyi's own sample prompts are long Chinese descriptions, and English works just as well. That covers the prompt language only; it isn't a promise of accurate Chinese lettering inside the image.
Yes, on paid plans: subscriptions include a full commercial license, and the Lifetime plan covers personal commercial use (see Pricing). The model weights are released under Apache 2.0; images made on the free plan are public.
Generate a run of Z-Image Base images, keep the direction that works and refine it from there.
