Z Image Turbo AI Image Generator

Every image in the groups below was generated on this page on 19 August 2026 — 14 prompts, 14 first-take results, 12 to 33 seconds each. Each group tests the claim above it, including the one it fails.

A photographic look, not a "free model" look

This is the strength reviewers keep circling: output that reads as a photograph rather than a render. The three below are single-sentence prompts, first take — no retries, no cherry-picking. Look at the weathered hands on the fisherman, the crust texture on the bread, the way the neon doubles in wet asphalt. For social content, mood boards and everyday images, this is usable as-is, and it cost nothing to find out.

3:2 · 1536×1024 · 12s

3:2 · 1536×1024 · 26s

3:2 · 1536×1024 · 23s

Three one-line prompts, three first takes, all under half a minute. None of these would embarrass a paid model — that is the whole argument for drafting here first.

Where it gives ground: briefs with exact specifics

These two prompts are copied word for word from our GPT Image 2 test set, where the flagship followed both to the letter. Z Image Turbo gets the scene, the light and the mood right — and misses the specifics. The counter holds four lemons where the brief says three, and the towel ends up under the fruit instead of between the objects. The three people are dressed exactly as asked, but the umbrella is open instead of closed, the apple sits in her right hand instead of her left, and the "full bodies" are cropped at the ankles. Read each brief beside its image and count for yourself.

The brief we gave it

First take, unedited

A kitchen counter scene containing precisely four items and nothing else: a cast-iron skillet at the back left, three whole lemons stacked in a small pyramid at the front right, a folded blue-striped tea towel between them, and a single sprig of rosemary lying diagonally across the towel. Soft window light from the left.

Three people standing in a row against a plain grey wall, exactly as described: on the left a woman in a mustard-yellow raincoat holding a closed black umbrella; in the centre a man in a navy turtleneck holding nothing; on the right a woman in a white shirt holding a single red apple in her left hand. Even frontal light, full bodies visible, no other people or objects.

Same briefs, two models: GPT Image 2 landed every constraint on its page; the free model landed the look and lost the specifics. That is the honest trade — and since it also ignores negative prompts by design, phrase everything as what you want, not what to avoid.

One free model, three output sizes

Z Image Turbo is free on every plan; what changes with your plan is the canvas. Free accounts render at a 512px base, Premium at 1024px, Ultimate at 1536px — the API's maximum. The three below are the same prompt with the same fixed seed, one per tier. Worth knowing before you rely on it: the model re-renders the scene at each size rather than upscaling, so composition shifts as the canvas grows — and in our test the 1536px version actually dropped the crackle glaze the prompt asked for while the two smaller ones kept it.

512×512 · 18s

1024×1024 · 17s

1536×1536 · 33s

Same prompt, same seed, only the output size changed. Note what the biggest render lost: the crackle pattern in the glaze. More pixels bought smoother ceramic, not more of what the prompt asked for — a genuine quirk of this model, not a tier upsell.

The free half of a paid workflow

Running Z-Image-Turbo yourself means a GPU with around 16GB of VRAM and a ComfyUI or diffusers setup; here it is a login. It is also the model our Ava agent reaches for by default when a free-plan brief is a straightforward text-to-image draft — and Ava routes past it the moment the brief calls for something it cannot do, sending typography work to GPT Image 2 and photoreal finals to Seedream. Every generation keeps its prompt and parameters in history, so the winning draft is one click away from a flagship re-render on the same subscription.

• Anyone trying AI images for the first time

  The generator on this page already has it selected, results land in under half a minute, and a miss costs nothing. There is no cheaper way to learn what prompts do.

• Social and content creators posting daily

  Nine ratios cover square feeds, 9:16 stories and 2:1 headers, and the photographic look above is usable as-is for everyday posts.

• Prompt writers iterating toward a paid render

  Draft ten variations here, keep the one that works, re-run it on GPT Image 2 or Seedream from history. The free drafts are what make the flagship credits go further.

• Poster makers working in English or Chinese

  Short display text — a headline, a date line, two hanzi — came back correct on the first take in both our tests. Keep it headline-sized and proofread anything longer.

• Not for: spec-tight briefs and fine print

  Exact counts, left-right assignments and dense typography drift, as the brief tests above show. When the image is a spec, that is what the flagships are for.

Credits, reference-image limits and ratio counts are read from our own model registry on 19 August 2026 — ratio counts exclude "auto", which is a setting rather than a ratio. Speeds are our own measurements with the sample size stated; cells we have not tested say so rather than borrowing a vendor number.