Alibaba Releases Qwen-Image-2.1 With Transparent Image Editing

Developers can now test one downloadable model for image generation, local edits, transparent images, and work guided by multiple references. Alibaba has released Qwen-Image-2.1, an open-weights image-generation and editing model with a 7-billion-parameter visual component.

Qwen-Image-2.1 is available through GitHub, Hugging Face, Model Scope, and a Hugging Face demo. Qwen says it can run on capable consumer GPUs such as an NVIDIA RTX 3090, giving developers a route to test the model outside a hosted product.

Transparency and reference images in the same release

The model natively generates and edits transparent RGBA images. Qwen says users can isolate objects or change text on transparent layers, a capability suited to image work where the background is not meant to be part of the finished asset.

It can also accept up to 10 reference images at once. Qwen identifies group portraits, virtual try-ons, and room design as examples of work that can draw on a supplied set of images rather than a text prompt alone.

Controls for a targeted edit

• Provide reference images alongside a generation or editing request.
• Mark the area to change with a circle, mask, or painted mark.
• Use those marks to guide a localized edit within the image.

A public release with a separate business decision

That licensing split makes the release more straightforward for research and evaluation than for a commercial product. Companies can inspect and test the available model, but a business deployment depends on securing different terms from Qwen.

Qwen also says architecture changes and key-value cache reuse speed inference, particularly when a request includes multiple reference images. Separately, it claims Qwen-Image-2.1 beats most closed models on its own benchmark; independent benchmark results were not available when the release was published.