Qwen Image 3.0

Overview

An image editing model that iteratively modifies images based on reference images, part of the Qwen-Image 3.0 series. Pass in 1–3 reference images and a positive prompt describing what to keep and what to change, and it can modify the scene, style, clothing, and on-image text while preserving key details of the original image, outputting 1K or 2K images. It is suited for developers who need to quickly transform existing assets into infographics, posters, UI concept drafts, or product showcase images, especially in content production scenarios with dense on-image text and high multilingual layout requirements.

Core Highlights

• Iterative editing with reference images, preserving original details — takes 1–3 images as input and modifies the scene, style, clothing, and text according to the prompt rather than fully redrawing, making it easy to iterate on the same asset repeatedly.
• Dense layout and small-font text rendering — designed for information-bearing visuals, with more stable layout and multilingual text rendering, suitable for text-dense outputs such as infographics and interface drafts.
• Prompts can describe complex layouts — the positive prompt supports up to 18,000 characters, enough to describe multi-element scenes, layout structures, and region-specific modification requirements.
• Controllable output specifications — supports 1K/2K resolution tiers, 8 aspect ratios, png or jpeg output, random seeds, and negative prompts, making results easy to integrate into production workflows.

Capabilities

Reference image editing: pass in 1–3 reference images along with a prompt to modify content while preserving key information from the original images, suitable for targeted edits to existing assets rather than regenerating from scratch.

Long prompts and negative constraints: the positive prompt supports long-text descriptions of layouts and areas to preserve or modify, while a negative prompt excludes unwanted elements for precise image control.

Layout and text rendering: can generate images with correct layouts and multilingual text, adapted to text-dense visual content production such as infographics and interface concept drafts.

Automatic prompt optimization: prompt_extend is enabled by default and can automatically expand prompts, reducing the description burden when input is brief; it can also be disabled to keep prompts as-is.

Output specifications and reproducibility: specify resolution tier, aspect ratio, and file format, and reproduce results via seed, making it convenient to batch-produce assets in a consistent style.

Use Cases

• Designers reworking existing posters or infographics — upload the original poster or layout draft, adjust copy, style, or layout via prompt, and obtain an updated visual asset ready for delivery.
• Product teams iterating on UI concept drafts — use existing interface screenshots as reference images, modify color schemes, layouts, or text, and produce high-fidelity interface concept images for review and prototype discussions.
• E-commerce operators updating product showcase images — use product images as references, adjust the scene, background, or model clothing, and produce product showcase images for detail pages and marketing materials.
• Multilingual material localization — use Chinese or English materials as reference images, replace the language and style, and generate promotional visual assets adapted to different markets.

Limitations and Selection Advice

Only image editing is supported; it does not have text-to-image capability, so pure text-to-image generation requires choosing the text-to-image route in the same series. Reference images are limited to 1–3, each no larger than 10 MB, and SVG is not supported, so complex multi-image composition scenarios will be constrained. The readability of small-font text is affected by layout density and language, so outputs should be checked one by one before use. This model is relatively new, and public independent evaluations are limited; it is recommended to validate results with a small number of samples before integrating it into batch workflows.