How to prompt Qwen3

Qwen Image 3.0 is Alibaba’s image model, and in the bot it is the one to reach for when the picture has to carry information: a labelled diagram, a poster with a headline, packaging with a brand name on it. It is the strongest model here at text, and the one that holds the longest brief. Both of those change how you write the prompt, because a model that reads every word rewards you for writing more of them.

The Qwen3 prompt formula

The same six slots as everywhere else in the bot cover most prompts: subject, action, setting, light and lens, framing, style. What differs is a seventh that only pays off here: the words in the picture, and where they sit.

A flat vector infographic on a warm off-white background explaining three coffee brewing methods, three equal columns each with a simple line-art icon above its heading, the headings reading “POUR OVER”, “FRENCH PRESS” and “ESPRESSO” in a bold geometric sans, one short line of muted grey text under each heading, a thin rule between the columns, the title “HOW TO BREW” centred across the top, generous margins, 4:3

Nothing there is decorative. Each clause is an instruction the model can act on: how many columns, what goes above the heading, what the headings say, how heavy the type is, where the title sits. Write a prompt that reads like a brief to a designer, not like a mood board.

Because the prompt rewriter is off, length works in your favour: nothing paraphrases your wording, so an extra clause about the margins or the rule between columns lands as written. The corollary is that vagueness is not filled in for you either. “Make it look professional” buys nothing; “generous margins, one accent colour, everything else greyscale” buys a layout.

Getting the lettering right

Qwen3 handles words better than anything else in the bot, which makes it worth learning where the ceiling actually is.

• Quote the exact string. the sign reads "OPEN DAILY 7–3" beats “a sign about opening hours”. Anything unquoted is treated as a description of the text, and you get plausible letters rather than yours.
• Name the surface. Printed on a paper label, painted on a brick wall, embossed on a matte box, cut into a neon tube. The surface decides the letterforms as much as the font description does.
• Say how big and where. “Filling the top third”, “small, bottom right, under the logo”. Unplaced text drifts to the middle and grows.
• Keep each string short. A headline plus a subhead plus three labels is comfortable. A paragraph of body copy is not: it comes back as convincing typographic texture that falls apart on reading.
• Describe the type, do not name a font. “A bold condensed grotesque, all caps, tight tracking” is followed. A font’s brand name usually is not.
• Re-run rather than rewrite when a single character is wrong. Lettering varies run to run, and the prompt that got you nine correct letters will often get you ten on the next pass.

Non-Latin scripts are where this model earns its reputation: Chinese, Japanese, Korean, Cyrillic and Arabic strings hold up here far better than on the other models. Quote them the same way, and keep them just as short.

Infographics and layout

An infographic fails in a predictable order: first the count is wrong, then the alignment, then the words. Write the prompt so it fixes them in that order.

Start by stating the structure and the count, before any styling: three equal columns, a five-step vertical timeline, a two-by-two grid of cards. Then say what one unit contains (icon, heading, one line of text), and the model repeats it across the rest instead of inventing a different arrangement per cell.

Then set the visual system: a background colour, one accent colour, whether the icons are line art or solid, whether there are rules between sections. A named system produces a consistent sheet; a pile of adjectives produces a collage.

Only then write the strings, quoted and placed. And ask for generous margins explicitly: without them the layout expands to the edge of the frame, which is the single most common reason an otherwise good result looks unusable.

Pick the ratio from the layout: 4:3 or 3:2 for a sheet, 9:16 for a story, 16:9 for a slide, 21:9 for a banner. There is no auto here, so this choice is always yours, and a layout prompt written for a wide frame does not survive being rendered tall.

Editing attached photos

Up to 3 photos go in one message with the instruction as its caption; the step-by-step guide covers the mechanics with screenshots. Address them by number in the prompt: keep the bottle from @image1, use the background of @image2.

Three is one fewer than the other models allow, and it is deliberate on the model’s side rather than ours: every attachment is priced, so send the references that have a job and drop the rest. State that job for each one. A reference with no stated role is absorbed into the mood instead of being used.

The edits this model is unusually good at are the ones involving text: swapping the words on a sign or a label while keeping the surface, the light and the perspective, restyling a poster around its existing headline, or putting a brand name onto packaging that already exists in your photo. Say what must not change (“same bottle, same label position, same studio light”) as explicitly as what must.

Qwen3 or Qwen3 Pro

Both tiers cost the same at 1K and 2K within their own family, which makes the workflow obvious: draft on Qwen3 at 2K for 2 credits until the structure and the strings are right, then re-run the same prompt on Pro. On a simple image with one short label the two look close enough that Pro is wasted credits. On a dense sheet with several headings, small labels and an icon set, Pro is where the small type stops smudging and the alignment stops drifting.

Common fixes

Ready-made prompts

The fastest way to learn a model is to run a prompt that already works and change one slot at a time. The prompts library has ready-to-paste prompts next to the images they produced, and the Qwen3 model page lists the resolutions, aspect ratios and credit costs.

Open the bot, pick Qwen3, choose a ratio, and paste one in.