Open-source Kokon 0.12 cuts image objects into layers and exports PSD, powered by Qwen-Image
Ok-Director4174 · reddit · 2026-11-05
Kokon, a free open-source desktop app built on Qwen-Image 2.1 Turbo (8 steps), shipped v0.11/0.12 with an experimental "Layers lab":
- Click objects to cut them onto separate layers; the same edit model repaints the background behind them so moves leave no holes. Move/scale/hide layers, Expand objects cropped by the frame, and export a PSD (Photoshop/GIMP/Photopea). "Blend into scene" redraws shadows and reflections after moves
- Honest limitations: removing large people sometimes paints a different person; crowds can merge or shatter; occluded regions aren't completed; blending all objects at once works best
- 0.12 also adds paint-by-color multi-edits, one-click Select, Prompt+ (Qwen3-VL encoder) for brushed areas, and EU AI Act metadata + invisible watermarking
- 0.11 added phone/LAN browser use, Docker, headless Linux, and a native Apple Silicon engine — about a minute per 1 MP image on M2 Pro, 29 s on M5 Pro using neural accelerators (4x faster)
Repainting behind one object takes 15–25 s on an RTX 5070 Ti. Runs fully locally, AGPL; Qwen weights are non-commercial.
Related event: Open-Source Kokon 0.12 Adds Click-Based Layer Extraction and PSD Export(2 posts)→
More from Multimodal
- Dev open-sources Ninada: 48 on-device 24kHz TTS voices across Sanskrit and 12 languages — init0 · 2026-11-06
- Self-taught mom's 9-minute AI fantasy film wins 4 awards at NY International Film Awards — minchoi · 2026-11-06
- ComfyUI Image Size Explained: Empty Latent, Aspect Ratio, Batch Size and Why 1080 Becomes 1088 — pixaromadesign · 2026-11-06
- AI-Generated 'WHATshake' Fast-Food Clip Amuses Reddit With Over-the-Top Character — OutillageFromage · 2026-11-06
- NitroFusion2 debuts: a 1-2 step text-to-image model built on Z-Image-Turbo — bdsqlsz · 2026-11-06
- Seedance 2 Video Shows Striking Wasteland Selfie With Realistic Dust and Ruins — Ok-Nerve941 · 2026-11-06