Introducing the image generation AI 'Qwen-Image,' capable of generating higher-quality images than OpenAI and Flux, and boasting astonishing text drawing capabilities that can naturally depict 'multiple lines of kanji characters.'
Qwen, Alibaba's AI development team, announced the image generation AI ' Qwen-Image ' on Monday, August 4, 2025. Qwen-Image excels at 'accurate text rendering,' a weakness of existing image generation AI, and can accurately render 'images containing multiple lines of Chinese text' and 'images containing both English and Chinese.' It also boasts high quality in general image generation and image editing.
Qwen-Image: Crafting with Native Text Rendering | Qwen
Qwen-Image is an image generation AI developed based on a technology called 'Multimodal Diffusion Transducer (MMDiT),' which uses separate weights for image and text representations, and is characterized by its high text drawing performance. Qwen-Image was instructed to depict a 'T-shirt printed with 'QWEN'' and a glass panel with the words 'Meet Qwen-Image – a powerful image foundation model capable of complex text rendering and precise image editing.' The images below are generated by instructing Qwen-Image to draw a glass panel with the words 'Meet Qwen-Image – a powerful image foundation model capable of complex text rendering and precise image editing.' It is possible to accurately draw even fairly long sentences and supports simultaneous drawing of English and Chinese.
Qwen-Image also allows you to precisely specify the position of text within an image, and can generate slide-like images like the one below.
The following figure compares the text rendering performance of 'Qwen-Image (blue),' 'GPT Image 1 [High] (green),' and 'Seedream 3.0 (light blue).' Qwen-Image showed the top score in Chinese rendering performance, and also outperformed GPT Image 1 [High] in some English rendering tests.
Qwen-Image also boasts high general image generation performance, with examples such as photo-like images, illustration-like images, and ink painting-like images being released.
It also allows you to perform editing tasks such as 'changing a character's pose,' 'changing the image style while maintaining the character,' and 'adding objects to an image' with high quality.
Below is a comparison of the image generation and image editing performance of 'Qwen-Image (blue)', 'GPT Image 1 [High] (light purple)', 'FLUX.1 Kontext [Pro] (light blue)', 'Seedream 3.0 (green)', 'FLUX.1 [Dev] (yellow)', and 'BAGEL (orange)'. Qwen-Image recorded scores that exceeded rival models in both generation and editing.
The model data for Qwen-Image is available at the following link:
Qwen/Qwen-Image · Hugging Face
https://huggingface.co/Qwen/Qwen-Image
Related Posts:
in Software, Posted by log1o_hf