The open-source Kokon 0.12 project has connected Qwen-Image 2.1 to a layered image-editing workflow: users can select an object to generate it as a separate layer, redraw the background behind occlusions, and export the result as a PSD. What matters most is that it does not stop at “generating a more photorealistic image”; it fills a practical gap that image-editing software has long left unfilled.
What this is
Kokon is a locally run image editor built around Qwen-Image 2.1 Turbo. The system first uses MobileSAM to identify a clicked object, then isolates the subject through a segmentation model and asks Qwen-Image to redraw the hidden background. After an object is moved, it can also attempt to recreate contact shadows and reflections. The developer provides a local server, Docker deployment, a Mac engine, and an MCP interface for agent-based control. Here, MCP—the standard for allowing external programs to invoke model tools—means automated workflows can directly generate or edit images.
The workflow remains experimental: deleting a large object may cause the model to invent another object, people who touch may merge or split apart, and fully occluded sections cannot be reconstructed. According to the project’s tests, background redraws take about 15–25 seconds per object on an RTX 5070 Ti and roughly one to two minutes on an M2 Pro.
Industry view
We believe this comes much closer to real professional needs than one-off text-to-image generation. What designers, e-commerce teams, and marketers actually need is often not a prettier standalone image, but an engineering workflow in which assets can continue to be modified, delivered, and reused. Separating objects into layers, preserving backgrounds, and exporting PSDs connect generative models to the existing software ecosystem instead of forcing users to learn an entirely new toolset.
The risks are equally clear: current results depend on repeated generation and manual selections, and the system’s speed and stability are not yet sufficient to replace traditional cutout and compositing software. More importantly, Qwen-Image uses a non-commercial license for its weights. Although Kokon itself is open source under the AGPL, enterprises using it commercially must review the model-weight and code licenses separately; “free and open source” cannot be treated as synonymous with “free for commercial use.”
Impact on regular people
For enterprise IT: Local execution and LAN services keep sensitive images off the cloud, making the system suitable for teams with strict asset-confidentiality requirements. GPU deployment, model licensing, and workflow stability, however, remain barriers to adoption.
For everyday work: People in non-design roles may also be able to perform basic edits such as changing product backgrounds or repositioning people quickly, reducing the time spent repeatedly rewriting prompts and manually compositing images.
For the consumer market: The free tool pushes the price of high-quality layered editing even lower, but in the near term these capabilities are more likely to appear as professional software plug-ins than to replace Photoshop.