What’s new
- From "good-looking" to "useful": the Qwen team frames each generation around one word — 1.0 was about "precision," 2.0 added "variety, completeness, beauty and authenticity," and 3.0 narrows it down to "real," aiming to make generated images an actual production tool.
- Prompt length up to roughly 4.5k tokens: enough room to describe a whole complex layout in one go — a newspaper page, a storyboard, an exam sheet.
- Legible text down to 10px: Qwen's own demos show it reproducing micro-level detail like pores and hair strands alongside that small text.
- Native rendering in 12 languages: Japanese, Korean, Spanish and more generate directly, and it can also mimic real interfaces — web pages, game UIs, livestream overlays.
What it’s good for
- Information-dense layouts: newspaper pages, storyboards, exam sheets, instruction manuals.
- Anything that needs small text to stay legible: simulated UI screenshots, annotated charts.
- Multilingual assets — not just Chinese and English, but native generation in 12 languages including Japanese, Korean and Spanish.
Prompting tips
- Describe the layout structure directly in the prompt: how many columns, where the headline sits, the image-to-text ratio — the more detail, the closer you get to real layout design.
- Don't be afraid of long prompts: the ~4.5k-token budget is enough to describe an entire layout's logic, so use the space.
- Write out any small text verbatim: if you need legible captions or labels, put the exact text in the prompt instead of leaving it to the model.
- Use negative prompts to cut unwanted elements: the API exposes a negative_prompt field for exactly this.
- Pick the right prompt-expansion mode: use "agent" mode when you want the model to add creative detail (text-to-image only), or "direct" mode — or turn it off — when you want your wording followed closely.
Front page layout for a tech newspaper: main headline "AI Is Reshaping Urban Transit" at the top,
a large photo with a caption on the left, three columns of body text on the right,
a data chart with labeled axes across the bottom, black-and-white newsprint texture,
headline type noticeably larger than body text.
Known limits
- Editing accepts at most 3 reference images, fewer than the 4 supported by the previous 2.0 line.
- The official announcement post doesn't disclose parameter count, benchmark scores or a technical report — less transparent than Alibaba Qwen's own continuously open-sourced Qwen-Image line (see the separate Qwen-Image entry).
- The "agent" prompt-expansion mode only works for text-to-image; it isn't available for editing.
Using it in Nomi
On the generation canvas, add a image card, pick “Qwen-Image 3.0” in the model picker and choose a mode (Text to image, Image editing). Add your references, write the prompt and generate. In the storyboard you can pick it per shot. The first time, connect any one of the providers above under Settings → Models.
