The Qwen team has open-sourced Qwen-Image-2.1, a 7B image model that combines text-to-image generation, transparent image creation, and image editing in one framework. Its visual generation module uses a 32-layer Single-Stream DiT architecture. Mixed-granularity attention and KV Cache reuse are designed to improve inference efficiency for multiple-image inputs while controlling GPU memory use. The model can generate standard images or transparent RGBA images from prompts, edit expressions and text while preserving transparency, and extract subjects from real photos into transparent layers. For editing, it accepts up to 10 reference images and supports region selection, brush-based editing, and independent masks. The article also highlights improvements in portrait and product consistency, text rendering, lighting, and detail across tasks such as panoramas, infographics, and storyboards. Qwen-Image-2.1 is available for free through GitHub, ModelScope, and Hugging Face.
