Kokon 0.12 introduces a Layers Lab feature enabling editable layer extraction from Qwen-Image 2.1 Turbo, a model lacking native layer output. The pipeline uses MobileSAM (ONNX, CPU, 44 MB) for object selection via user clicks, generating masks to isolate objects onto individual layers with background inpainting. Users can manipulate layers locally or export them as PSD files, implemented on stable-diffusion.cpp and a custom Mac engine.