[Qwen Open Sources Qwen-Image-2.1, 7B Unified Image Generation and Editing with Transparent Image Support] According to DeepFlow TechFlow, on September 20, the Qwen team open-sourced Qwen-Image-2.1, integrating text-to-image generation and image editing into a single model. The visual generation module features only 7B parameters and natively supports the generation and editing of transparent images. The model supports up to 10 reference images and allows localized editing through circular selection, doodling, or independent masks. Architecturally, the model employs mixed-granularity attention and reuses KV cache to enhance inference efficiency, while improving text rendering, portrait lighting, and fidelity in character and product representation. It covers tasks such as panoramic images, infographics, and storyboarding.