The Qwen team at Alibaba has open-sourced Qwen-Image-2.1, an image model that combines text-to-image generation and image editing in a single system. According to the company, the model is designed to balance generation quality, inference efficiency, and cost, with a visual generation component of just 7 billion parameters.
The release highlights four main improvements: a compact and efficient architecture, native transparency with unified creation and editing, versatile editing capabilities, and more realistic textures and refined aesthetics. The model can generate both regular and transparent images from prompts, edit transparent layers, and extract subjects from photographs as RGBA layers.
On the editing side, Qwen-Image-2.1 accepts up to 10 reference images, supports local edits specified by circles, painted annotations, or separate masks, and aims to preserve the identity of people and the defining features of products such as text, textures, and shape. The company also demonstrates task coverage including panorama, infographic, and storyboard generation.
Efficiency gains come from a mixed-granularity attention architecture: text, including the system prefix and editing instructions, uses a token-level causal mask, while image generation uses a chunk-level mask. KV cache reuse lets input images and editing instructions serve as static context that is computed and cached in the first step, reducing memory usage and improving inference speed.
Qwen says the model's visual generation component contains 32 Single-Stream DiT layers. The team positions Qwen-Image-2.1 as a practical tool for design, content creation, e-commerce, and visual storytelling, and points to the Qwen-Image-Bench comparison for performance against other open-source and closed-source models.