Qwen Image
Qwen's 20B-parameter MMDiT text-to-image model. It draws legible English and Chinese text into the picture, from a single shop sign to a multi-line poster.
Overview
Qwen Image is the 20B-parameter MMDiT image foundation model that Qwen released on August 4, 2025. This page runs it as text-to-image: you write a description and get back one or more pictures.
Qwen leads its release material with text rendering. Its examples include multi-line layouts and a full handwritten paragraph. They are in English or Chinese, and in one of them a handwritten glass board carries an English sentence followed by a Chinese one. The model card says text rendering is especially strong for Chinese.
Away from lettering, Qwen shows the model working in photorealism, impressionist painting, anime and minimalist design. The samples on this page are all photographic: a firefighter carrying a child away from a fire, a climber hanging off a cliff at sunset, a drift car throwing sparks past a tyre wall, a surfer inside a breaking wave.
What you can build
- Posters and banners with the headline drawn into the image itself.
- Menu boards and packaging mockups with English and Chinese copy side by side.
- Slides and infographics with a title and labelled sections, the kind of layout Qwen uses in its own demos.
- Storyboard frames and editorial illustrations written as scene descriptions.
- Up to 8 variations of one prompt per request, for a designer to pick from.
Inputs
- A text prompt. Qwen's examples put the exact words to render in quotation marks, such as the text on a chalkboard sign or a neon light.
- An optional negative prompt for things to keep out of the image.
- Width and height, each up to 2048 pixels, with 1024 by 1024 as the default.
- Inference steps from 1 to 500 (default 25) and a guidance scale from 1 to 20 (default 3.5).
- The number of images, from 1 to 8.
- A seed. Leave it at 0 for a random result, or fix a number to repeat one.
Outputs
One image per requested sample, at the width and height you set. The samples here include 1024 by 576 widescreen frames and 1024 by 1024 squares.
Recommended settings
Qwen's quick-start example uses 50 inference steps, twice this page's default, with a CFG scale of 4.0. It also gives a width and height for each aspect ratio, all inside the 2048-pixel limit:
- 1:1 at 1328 by 1328
- 16:9 at 1664 by 928, and 9:16 at 928 by 1664
- 4:3 at 1472 by 1140, and 3:4 at 1140 by 1472
- 3:2 at 1584 by 1056, and 2:3 at 1056 by 1584
For English prompts, the same example adds ", Ultra HD, 4K, cinematic composition." to the end of the prompt.
Limitations
- The model card lists only English and Chinese, and Qwen's published text examples stay within those two languages.
- Qwen also presents editing work for Qwen Image, for example style transfer or changing the text inside an existing picture. That needs an input image, and this page takes text only.
Safety & compliance
Qwen Image is released under the Apache 2.0 license.
Example prompts
Great starting points for Qwen-Image.
API quick start
Run Qwen-Image with a single API call.
{
"prompt": "A mountain climber hangs from a cliff edg…",
"negativePrompt": "...",
"steps": 25,
"scale": 3.5
}curl -X POST "https://api.wiro.ai/v1/Run/qwen/qwen-image" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "A mountain climber hangs from a cliff edg…",
"negativePrompt": "...",
"steps": 25,
"scale": 3.5
}
JSONOther models you might like
Explore similar image generation models.



