MODEL IDalibaba:qwen-image@3.0
live

Qwen-Image-3.0

Alibaba
by Alibaba

Qwen-Image-3.0 is a unified image model from Alibaba for both text-to-image generation and prompt-guided image editing. It supports one to three reference images for editing, automatic prompt expansion, negative prompts, reproducible seeded generation, and adaptive or user-specified output sizes. It supports text-to-image resolutions from 512x512 up to 2048x2048 and image editing resolutions up to 1440x1440, making it a strong fit for high-quality visual creation, controlled edits, and multi-image reference workflows.

Qwen-Image-3.0

Dense layouts and information design

How to generate infographics, comparison grids, newspaper pages, and interface mockups with Qwen Image 3.0 by spending its long prompt budget on every panel.

Image models fall apart on information design. A diagram with eight labels comes back with four of them misspelled and two pointing at nothing, so anything with real structure gets built in a design tool instead. Qwen Image 3.0 takes instructions of around 4,500 tokens and renders what they describe, which makes a fully specified infographic a single request rather than a starting point.

The title, the four numbered callouts, the statistics strip, and the footer were all written into one prompt. Nothing in that poster was composited afterwards.

Set promptExtend to false for every request in this guide. A layout spec is exactly the kind of prompt the rewrite will reword, and rewording a spec means losing panels and figures you carefully placed.

Request shape

Nothing here needs special parameters. The layout lives in positivePrompt, and the only other thing that matters is spending enough pixels for the small type to survive:

import { createClient } from '@runware/sdk'

const client = await createClient({ apiKey: process.env.RUNWARE_API_KEY })
await client.connect()

const [result] = await client.run({
  model: 'alibaba:qwen-image@3.0',
  positivePrompt: 'A single-page explainer poster titled "HOW A HEAT PUMP WORKS". The title sits across the top in large bold condensed capitals. The centre carries a technical cutaway diagram of a house with an outdoor unit on the left and an indoor unit on the right, connected by a loop of refrigerant pipework with directional arrows. Four numbered circular callouts sit around the loop, each with a short label beside it. Below the diagram, a horizontal strip of three statistics in large figures with small captions underneath. Flat editorial infographic design, crisp small type, no photographic elements.',
  width: 2176,
  height: 3008,
  providerSettings: {
    alibaba: {
      promptExtend: false
    }
  }
})
import asyncio
import os

from runware import Runware


async def main():
    async with Runware(api_key=os.environ["RUNWARE_API_KEY"]) as client:
        results = await client.run({
            "model": "alibaba:qwen-image@3.0",
            "positivePrompt": "A single-page explainer poster titled \"HOW A HEAT PUMP WORKS\". The title sits across the top in large bold condensed capitals. The centre carries a technical cutaway diagram of a house with an outdoor unit on the left and an indoor unit on the right, connected by a loop of refrigerant pipework with directional arrows. Four numbered circular callouts sit around the loop, each with a short label beside it. Below the diagram, a horizontal strip of three statistics in large figures with small captions underneath. Flat editorial infographic design, crisp small type, no photographic elements.",
            "width": 2176,
            "height": 3008,
            "providerSettings": {
                "alibaba": {
                    "promptExtend": False
                }
            }
        })


asyncio.run(main())
curl https://api.runware.ai/v1 \
  -H "Authorization: Bearer $RUNWARE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '[
    {
      "taskType": "imageInference",
      "taskUUID": "f7a8b9c0-1d23-4e45-8f67-9a0b1c2d3e45",
      "model": "alibaba:qwen-image@3.0",
      "positivePrompt": "A single-page explainer poster titled \"HOW A HEAT PUMP WORKS\". The title sits across the top in large bold condensed capitals. The centre carries a technical cutaway diagram of a house with an outdoor unit on the left and an indoor unit on the right, connected by a loop of refrigerant pipework with directional arrows. Four numbered circular callouts sit around the loop, each with a short label beside it. Below the diagram, a horizontal strip of three statistics in large figures with small captions underneath. Flat editorial infographic design, crisp small type, no photographic elements.",
      "width": 2176,
      "height": 3008,
      "providerSettings": {
        "alibaba": {
          "promptExtend": false
        }
      }
    }
  ]'
runware run alibaba:qwen-image@3.0 \
  positivePrompt="A single-page explainer poster titled \"HOW A HEAT PUMP WORKS\". The title sits across the top in large bold condensed capitals. The centre carries a technical cutaway diagram of a house with an outdoor unit on the left and an indoor unit on the right, connected by a loop of refrigerant pipework with directional arrows. Four numbered circular callouts sit around the loop, each with a short label beside it. Below the diagram, a horizontal strip of three statistics in large figures with small captions underneath. Flat editorial infographic design, crisp small type, no photographic elements." \
  width=2176 \
  height=3008 \
  providerSettings.alibaba.promptExtend=false
{
  "taskType": "imageInference",
  "taskUUID": "f7a8b9c0-1d23-4e45-8f67-9a0b1c2d3e45",
  "model": "alibaba:qwen-image@3.0",
  "positivePrompt": "A single-page explainer poster titled \"HOW A HEAT PUMP WORKS\". The title sits across the top in large bold condensed capitals. The centre carries a technical cutaway diagram of a house with an outdoor unit on the left and an indoor unit on the right, connected by a loop of refrigerant pipework with directional arrows. Four numbered circular callouts sit around the loop, each with a short label beside it. Below the diagram, a horizontal strip of three statistics in large figures with small captions underneath. Flat editorial infographic design, crisp small type, no photographic elements.",
  "width": 2176,
  "height": 3008,
  "providerSettings": {
    "alibaba": {
      "promptExtend": false
    }
  }
}
Response
[
  {
    "taskType": "imageInference",
    "taskUUID": "f7a8b9c0-1d23-4e45-8f67-9a0b1c2d3e45",
    "imageUUID": "4b7c1e58-0a92-4d36-b1f4-7e2a5c8d0f39",
    "imageURL": "https://im.runware.ai/image/os/a14d18/ws/2/ii/4b7c1e58-0a92-4d36-b1f4-7e2a5c8d0f39.jpg"
  }
]

Spending the prompt budget

Most prompting advice tells you to be concise. That advice is wrong for this model. The instruction window runs to roughly 4,500 tokens, which is several hundred words, and information design is where the whole budget gets used. Both sheets below are the same size and the same request, one described in a sentence and one specified in full:

The short prompt produces something that looks like an infographic and says nothing you chose. The long one carries twenty data values, and every one of them is a value you can check against your own source. The prompt is the content, so a layout you cannot describe is a layout you cannot generate.

Write these prompts the way you would brief a designer. Name the sheet, then the title block, then each region in reading order, quoting the copy for each. Working top to bottom is what keeps the model's placement aligned with your intent.

Multi-panel grids

A single request can hold several unrelated panels. The sheet below is four separate reference graphics in one pass, each with its own heading, illustration, and labels:

Four unrelated subjects sit side by side without bleeding into each other. The pattern that makes this work is addressing cells by position: "top left cell, headed", "bottom right cell, headed". Describing the grid once and then listing contents leaves the model to guess the mapping, and a stated position per cell removes the guess.

Declaring a shared visual treatment once, at the top, is what keeps four panels reading as one sheet rather than four clippings.

Editorial pages

The same approach builds a full page of editorial layout, where the structure carries as much meaning as the copy:

Body copy is the one place to stop specifying. Quote the masthead, the headline, the deck, and the caption, then let the columns fill themselves, because nobody reads the body text of a mock front page and specifying it would spend most of your budget on it. Name the structure instead: how many columns, where the subheading falls, where the sidebar sits.

Interface mockups

Product screenshots for landing pages and decks are the same problem wearing different clothes. Values, labels, and navigation all have to be legible and plausible:

Every figure on that screen was chosen rather than invented, which is the difference between a mockup you can ship on a marketing page and one that quietly claims a 900% conversion rate. Write the numbers you want to be seen making, and give the interface a single accent colour so the result reads as a designed product rather than a generic template.

Dense tables and small type

The hardest case is a page that is mostly small text in a rigid structure:

Twelve label and value pairs in a fixed order, plus units and a footnote. Listing the rows in the order they appear is what holds the sequence, since a table described as a set of facts comes back in whatever order the model prefers.

Qwen documents legible rendering down to roughly 10px type, which is the constraint that decides your canvas. A spec sheet at 1024 × 1024 puts its rows below that floor and the values turn to texture, so dense pages want the top of the pixel budget: 2176 × 3008 portrait or 3200 × 2048 landscape, both just under the 6,553,600 ceiling.

Tips

  1. Write the longest prompt of your life. The budget is around 4,500 tokens and a dense sheet needs most of it. Brevity costs you control here rather than buying clarity.

  2. Describe regions in reading order. Title block, then each panel top to bottom, then the footer. The order you write in is the order the model lays out.

  3. Address panels by position. "Top left cell" and "lower right corner" bind content to place. A list of contents without positions gets shuffled.

  4. Quote every value that has to be right. Figures, units, and percentages are the reason the asset exists, and anything unquoted is invented.

  5. Skip the body copy. Specify headlines, labels, and captions, then let filler text fill itself and spend the budget on structure.

  6. Render at the top of the pixel budget. Small type is the first thing to break, and dense layouts are almost all small type.

  7. Keep promptExtend off. A rewrite applied to a layout spec drops panels and reworks the copy you were most careful about.