How to use Z-Image for AI image generation

Z-Image (model name z-image-turbo) is a single-image, text-to-image model — you give it one written prompt and it generates one PNG. There is no reference image, no multi-image composition, and no video: it's the fast, focused tool for turning a description into a still picture.

Its two standout traits are speed and text rendering. It's built to generate quickly rather than to maximize every last bit of fidelity, and it reproduces both English and Chinese text inside an image more reliably than most general-purpose models — useful for posters, signage, packaging mockups, or any shot where legible in-image text matters.

Lead with the subject and composition in concrete terms — who or what is in frame, their pose, and where the camera sits ("full-body portrait, centered composition, direct eye contact with the camera"). Concrete framing language gives the model a clear target instead of a vague mood.

Layer in style and setting after the subject: describe the background, lighting, and color palette as their own clauses. A prompt that moves from subject → outfit/detail → background → lighting → framing tends to render more coherently than one long unordered sentence.

If you need text to appear in the image — a sign, a label, a piece of dialogue in a speech bubble — write it out exactly, in quotes, and say where it appears. Z-Image's text rendering is one of its strengths; give it a precise string rather than a paraphrase of what the text should say.

prompt_extend runs your prompt through an LLM rewriting pass before generation, returning the optimized prompt and its reasoning alongside the image. It helps most when your original prompt is short or vague — the rewrite adds detail you didn't specify, which can turn a thin idea into a fuller image.

Leave it off when you've already written a detailed, deliberate prompt — like the layered subject → style → background → lighting structure above. At that point a rewrite is more likely to drift from your exact intent than improve it, and it always costs more and takes longer to return.

Image looks generic or doesn't match what you imagined: the prompt described a mood rather than a concrete scene. Name the specific subject, setting, colors, and framing you want present, in that order.

Text in the image is garbled or missing: the prompt paraphrased the text instead of quoting it exactly, or asked for more than one text element without saying where each belongs. Quote the exact string and state its placement.

Output looks stretched or oddly cropped: the resolution isn't one of the recommended pairings for that aspect ratio. Switch to a listed size.

Result is close but not quite right: change only the seed and re-run before touching the prompt — generation is probabilistic, and a different seed alone often fixes a near-miss.

Request returns a moderation error: both the prompt and the output image are checked for policy compliance. Rework the prompt to remove the flagged content rather than resubmitting unchanged.

Reach for Z-Image when you want a single, fast still image from a text description — it's lighter and quicker than models built for maximum photorealism, which makes it a strong fit for drafts, concept passes, and any image with an in-picture text element like a sign or label.

If your shot needs to combine several separate reference photos into one scene, that's a job for a reference-based generator instead. If you're aiming to animate a finished image afterward, a Z-Image output makes a clean first-frame source for an image-to-video model — generate the still here, then hand it off for motion.