Qwen Image 3

Text-to-image creates a new picture from a prompt alone, while Edit starts from one to three reference images. Use text-to-image for posters, infographics, UI mockups and storyboards. Use Edit to change text, swap clothing or objects, or composite subjects while keeping identity and untouched regions.

Use the Edit variant. Attach the image and describe the new text plus what must stay unchanged, for example 'preserve the building and lighting, keep everything else unchanged'. The frame is re-rendered, so some unrequested drift can occur, and output size is limited by the pixel budget when references are present.

It is best at images that carry readable words or structure: infographics, bilingual Chinese and English posters, UI mockups, storyboard grids, labeled product diagrams and identity-preserving edits. The look is clean, editorial and evenly lit, closer to a designed document than a painting.

Chinese and English are its strongest languages, and 2K resolution is recommended for dense captions and small text. Korean and Japanese can show occasional character errors, so proofread them. Quote the exact text in your prompt and say where it goes and in what style.

Skip it for painterly or stylized fine art, cinematic portraits where facial beauty matters most, and prompts that need an exact count of items or precise geography. Vector-clean logos are also a weak spot.

It was created by Alibaba's Qwen team, with a release date of 2026-07-21. It's a unified text-to-image and instruction-editing model that switches to edit mode when you attach one to three reference images.