Alibaba's Qwen-Image-Edit-2509 Merges Multi-Image Editing and ControlNet Into One Open Model
Alibaba's Qwen team ships a monthly refresh that adds multi-image composition, ControlNet conditioning, and much stronger identity preservation.
Takeaways
- Qwen team released Qwen-Image-Edit-2509, a 20B parameter open-source image editor under Apache 2.0.
- New multi-image editing supports person+person, person+product, and person+scene compositions with 1 to 3 inputs.
- Consistency upgrades preserve faces across pose and style changes, plus product identity for posters.
- Text editing now covers font type, color, and material, not just content swaps.
- Native ControlNet support for depth, edge, and keypoint conditioning inside the same pipeline.
- Available via Qwen Chat, the GitHub repo, and Diffusers as
QwenImageEditPlusPipeline.
Alibaba's Qwen team just pushed the September iteration of its open-source image editor, and it is a substantial jump rather than a routine tune-up. Qwen-Image-Edit-2509 keeps the same 20B parameter backbone as the August release but adds the three things people kept asking for: multi-image inputs, tighter identity consistency, and native ControlNet conditioning. It ships under Apache 2.0, runs on Hugging Face, and has already crossed 455K downloads.
From single-image tweaks to scene composition
The headline feature is multi-image editing. Building on the original Qwen-Image-Edit architecture, the model was further trained via image concatenation to handle combinations such as "person + person," "person + product," and "person + scene." Quality holds up best with one to three input images.
In practice you can drop in a person photo plus a product shot and ask the model to fuse them into a coherent scene, instead of chaining several inpaint passes together. Consistency improvements make the results trustworthy for identity-sensitive edits like portraits or branded products, so the workflow shifts from tweaking one picture at a time to composing whole scenes in a single call.