Qwen-Image: Image Creation with Native Text Rendering

Qwen-Image has been released as an image foundation model based on 20B MMDiT. Its core strengths are the ability to naturally render complex sentences and layouts, and the precision to maintain both semantics and visual realism during editing.

The main features are as follows.

• Superior Text Rendering: Generates alphabetic scripts like English as well as ideographic scripts like Chinese with high accuracy, including multi-line layouts and paragraph-level text.
• Consistent Image Editing: Expanded multi-task training improves consistency in editing tasks such as style transfer, addition/removal, detail enhancement, text editing, and pose adjustment.
• Strong Cross-Benchmark Performance: Shows performance surpassing existing models across general generation and editing benchmarks.

The evaluation results are also strong. GenEval, DPG, and OneIG-Bench verified general image generation performance, while GEdit, ImgEdit, and GSO confirmed editing capability. In text rendering, it shows particular strength on LongText-Bench, ChineseWord, and TextCraft, demonstrating a notable advantage in Chinese text generation.

The demos showcase a variety of cases, including Chinese signage, couplets, long paragraphs, bilingual phrases, posters, and PPTs. It naturally arranges everything from short text to long descriptive passages, switches between English and Chinese within a single scene, and reproduces sentences on small pieces of paper or handwriting on glass with relative accuracy.

With its strength in text, its range of applications is broad. It presents a direction toward becoming a general-purpose image model that covers content creation such as posters, slides, and brand mockups, as well as general artistic style generation and image editing.