Kokon wrangles editable layers out of edit-only Qwen-Image 2.1: pipeline, wins and honest failures
Ok-Director4174 · reddit · 2026-11-05
The developer of Kokon, a local photo editor built on Qwen-Image 2.1 Turbo, released 0.12 with a Layers lab: click objects to get each on its own layer with the background repainted, move them, export PSD. The pipeline: MobileSAM (44 MB, CPU) for masks, a CPU matting model or Qwen edge-redraw for cutouts, a masked Qwen edit for background fill (15-25s per object on RTX 5070 Ti, 1-2 min on M2 Pro), and an optional blend pass for contact shadows. Honest failure list: large-object removal invents new objects, touching people merge/split, occluded parts are never completed. Also new: a local server mode with LAN access plus an MCP server, and a custom Swift/Metal engine for Apple Silicon (4x faster; M5 neural accelerator path cuts 37s to 29s). Free, AGPL, NVIDIA 12 GB, Qwen non-commercial weights.
Related event: Open-Source Kokon 0.12 Adds Click-Based Layer Extraction and PSD Export(2 posts)→
More from coding & agent
- Team's 24/7 infra-monitoring agent spots anomalies 'before the fire,' user says — antoine_chaffin · 2026-11-06
- Delvetown: a human-agent "post-scarcity" web community, plus agents emailing researchers unbidden — Miles_Brundage · 2026-11-06
- Every team demos Codex and Claude workflows for running AI as your OS — every · 2026-11-06
- Hyperagent launches Sizzle, an agent that reads PRs and auto-produces polished demo videos — aakashgupta · 2026-11-06
- Vercel publishes CLI design guidelines for humans and AI agents — JohnPhamous · 2026-11-06
- Dev builds a sculpting tool for Apple Vision Pro using Codex — OpenAIDevs · 2026-11-06