Qwen Image 3 Pro AI Image Generator

Every image below was generated on this page on 19 August 2026 — ten prompts, ten first-take results, 57 to 93 seconds each, with the exact prompt shown. Including the test that exposed a real limit.

Chinese typography it sets like a native

This is the home ground of the Qwen image family. We asked for a concert poster with a four-character Chinese title running vertically down the left in a songti face, and got exactly that: every character correct, the stroke weight consistent, the English subtitle spelled right in spaced capitals — and a red seal chop by the title that was not in the brief, which is precisely what a designer trained on Chinese poster convention would add. The bilingual tea tin further down this page came back with both scripts correct too. If your work switches between English and Chinese in one frame, this is the strongest reason to pick this model.

Three type sizes on one board, every word correct

A menu board is a harder text test than a headline: three sections, six dish names, six prices and a small footer all have to hold in one frame. Ours came back with every word and every price correct, from the hand-painted heading down to 'Served till 11am'. One quirk worth knowing: our brief separated dishes with slashes, and the model painted those slashes onto the board as literal marks. It follows text instructions almost too literally — do not put punctuation in quotes unless you want it rendered.

Where the small-print claim breaks: a full paragraph

Alibaba's launch claims include legible text at 10 pixels. So we asked for a vintage recipe card carrying a full forty-word method paragraph in small type. The result is the most honest image on this page: the title and the first five lines are real, readable English, then the middle of the paragraph collapses into ink-shaped marks that only look like words, before recovering cleanly for the final seven lines. Two of the legible words are also misspelled — 'soffened' and 'toothpack'. Display-size text is reliable here; long small print is not. Keep paragraphs off the image, or set them yourself afterwards.

The same brief at 1K and 2K

The Pro tier prices 1K and 2K separately, so we ran the identical prompt once at each. Both frames keep genuinely weathered skin — creases, stubble, sun damage — rather than the smoothed faces this model family gets criticised for, and the 2K frame resolves visibly finer net and rope detail. Both runs also did something we never asked for: they wrote signage into the scene, 'PIER 7' on a mooring post in one and 'HARBOR FISHERIES' on a crate in the other, correct at display size while the smaller line beneath the crate stencil degraded. That is this model's character in one pair: it volunteers text constantly, and the size threshold decides whether the text survives.

1K · 3:2 · 76 seconds

2K · 3:2 · 93 seconds

Same prompt, both first take. We did not fix the seed, so the compositions differ — what is comparable is the rendering, not the framing. For a like-for-like variation workflow, fix the seed in the advanced controls.

We probed the reported chart weakness — this run passed

The sharpest criticism in independent reviews is data integrity: charts that look right while being numerically wrong. We probed it directly — a bar chart brief with five exact values and strict proportionality. The result printed all five numbers correctly, kept Thursday tallest and Friday shortest, and then went beyond the brief: it added a 'Total Cups Sold: 950' line, which is the correct sum of our five values, and peak-day callouts that match the data. The one artifact is a legend entry for a 'Daily Target' series that appears nowhere in the chart. One passed probe does not retire the criticism — the reviewers' failures are documented — but it does mean the failure is not guaranteed. Check every number before a chart ships; on this image that check takes a minute.

• Bilingual brands and cross-border sellers

  Packaging, posters and store assets that carry Chinese and English in the same frame, both set correctly — the tea tin and guqin poster above are exactly this job. Have a native reader proofread anything customer-facing.

• Menu boards, signage and price lists

  Three type sizes, six prices and a footer line all held in our menu test. Keep the smallest text at label size, not paragraph size, and it stays readable.

• Detail-heavy product and editorial shots

  The watch-movement macro and the noodle tray show what the Pro tier buys: coherent small structure — gear teeth, steam, scallions — instead of smoothed-over texture.

• Infographic and layout drafts

  Our five-value chart probe came back numerically correct, including a self-added total. Treat every number as unverified until you have read it — the failure mode reviewers describe is real, it just is not guaranteed.

• Not for: paragraphs on the image, or scripts you cannot read

  Our forty-word paragraph collapsed mid-way, and independent testers report vowel-level errors in Korean. Set long copy yourself, and proofread any script you cannot verify by eye.

Credits, reference-image limits, resolutions and ratio counts are read from our own model registry on 19 August 2026 — GPT Image 2 prices on a resolution × quality matrix, which is why its cell is a range rather than one number. Speeds are our own measurements with the sample size stated, taken on each model's own page. Cells we have not tested say so rather than borrowing a vendor number.