Qwen Image 3 AI Image Generator

Every image below was generated on this page on 19 August 2026, with the exact prompt shown — including the chart test it failed. Nothing was retouched or picked from a batch.

Menus, posters and price lists that come back spelled right

This is the model's home ground. We gave it a printed diner menu with three sections and seven prices at small type sizes, an event poster with a headline, a schedule line and a small footer, and a bookstore poster with a vertical Chinese title over an English subtitle. All of it came back correctly spelled and correctly set — including things we never asked for: the menu grew a plausible "Est. 1952" and the poster a whole sidebar of extra labels, all spelled right, because automatic prompt rewriting is on by default and enriches short briefs. If you want only your text and nothing else, say so in the prompt.

2K · 2:3 · 113s

2K · 3:4 · 105s

2K · 2:3 · 95s

The test is not the headline — headlines are easy. It is the 7.50 and the 6.80 holding at menu-item size, and both scripts landing in one workspace. One caveat from the same batch: the model happily adds correctly-spelled content you did not ask for, so brief it tightly when the layout is fixed.

Attach a photo, keep the label, change the world around it

The same model edits. We generated a product shot of a labelled tea caddy, handed it back as a reference, and asked for the identical tin at an outdoor winter market at dusk. The tin, its shape, its colour and both lines of label text survived the move; the lighting was rebuilt for the new scene, down to a warm highlight on the metal against cold blue snow light. Up to 3 reference images per run — enough for product-in-scene work, not for the many-reference composites the Nano Banana channels are built for.

Prompt

Output image

Studio product photograph of a matte sage-green tea caddy on pale travertine, its label reading 'KIYOMI' in a fine serif with a smaller line 'ceremonial grade matcha · 40g' beneath, soft directional daylight from the right, one clean specular edge highlight, shallow depth of field, e-commerce hero framing.

Keep the same sage-green tea caddy from the reference image exactly as it is — same label text 'KIYOMI', same shape and colour — but place it on a weathered wooden stall at an outdoor winter market at dusk, light snow falling, warm string lights bokeh behind, cold blue ambient light with one warm highlight on the tin.

One honest nit from step 1: the label text is correct but its line hierarchy shifted — "ceremonial" moved up beside the brand name instead of leading the second line. The edit then preserved that layout faithfully, wrong line break and all.

The pricing quirk: 2K costs exactly what 1K costs

This is where the standard tier genuinely differs from its Pro sibling. On Qwen Image 3 Pro, 2K costs more than 1K; here they are priced identically, so this page defaults to 2K and the only thing 1K buys you is a faster loop. We ran the same typewriter macro at both tiers: 57 seconds at 1K, 84 at 2K, with roughly four times the pixels at 2K. Iterate at 1K if you are impatient; switch to 2K for anything you keep.

1216×832 · 57s

2496×1664 · 84s

Same prompt, same settings, only the resolution tier changed. The cost is the same either way — which is the whole point of this model, and the reason 2K is the default here.

The chart test — where we stop trusting it

Independent testers reported that its data charts look right and are wrong, so we ran our own: a five-bar chart with values we supplied. Every label and every number printed correctly — and the geometry underneath them is broken. The axis repeats one value and puts another out of order, the tallest bar tops out below where its printed value says it should, and the footer credits a "Global Coffee Report 2025" that does not exist, for data we invented. Treat it as a typographer, not an analyst: it will set your chart title beautifully, but build real charts in a charting tool.

2K · 3:4 · 118s

Look closely at the left axis: 1.5 appears twice and 2.0 sits above 2.5. The prompt asked for bars "strictly proportional" to their numbers; they are not. A page that only printed the successes would not tell you this.

• Cafés, restaurants and local businesses

  Menus, price boards and event flyers where every price has to be right on the first take. Our seven-price menu test is exactly this job, and it passed.

• Marketers working in English and Chinese

  One workspace for both scripts — our bookstore poster set a vertical Chinese title and an English subtitle in the same frame, both correct. Proofread any script you cannot read yourself; reviewers caught Korean errors.

• Anyone iterating in volume

  A flat price whatever the resolution, so exploring ten directions costs what it costs — no tier arithmetic. Draft at 1K for speed if you like, keep the finals at 2K for free.

• Product teams reusing existing shots

  Attach up to 3 references and move a product into a new scene with its label text intact, as in our winter-market edit. For composites with many references, use a Nano Banana channel instead.

• Not for: charts carrying real data

  Our chart test printed the right numbers on broken geometry and invented a source. Set real data in a charting tool, then let this model do the poster around it.

Credits, reference-image limits, resolutions and ratio counts are read from our own model registry on 19 August 2026 — the GPT Image 2 ratio count excludes "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.