← sources

GPT-Image-style photo editing on your own GPU (12 GB+): I built a zero-setup app for Qwen-Image 2.1 on stable-diffusion.cpp

source: reddit (reddit:LocalLLaMA)published 2026-10-14 17:10ingested 2026-10-15 02:42

retained text

Every local setup I tried for Qwen-Image 2.1 meant a Python env, a node graph and an evening of picking quantization, offloading and resolution for my card. **Kokon** is a *free, open-source desktop app* that does all of that for you and gives you a plain photo editor on top.

* **Zero setup:** first run detects your GPU and memory and downloads the matching GGUFs (Qwen-Image 2.1 Turbo Q4\_K + Qwen3-VL 8B text encoder, \~12 GB, checksummed). Each job is sized automatically: working resolution, VRAM/RAM split, tiled VAE, retry if it runs out of memory.
* **Engine:** stable-diffusion.cpp (ggml), so no Python, no PyTorch. CUDA, Vulkan and Metal builds.
* **An editor, not a workflow:** type the change, or brush an area with Change / Remove / Replace / Add. The brushed area is marked on a second reference image, so the model actually sees it. Versions with undo, before/after, upscaler comparison.
* **Hardware:** NVIDIA 12 GB+ (CUDA). Apple Silicon (Metal, beta, slow: \~4 min per image on a 16 GB M2 Pro). AMD / Intel Arc 12 GB+ via Vulkan (beta, only tested on NVIDIA so far, reports very welcome).
* **Speed:** 5070 Ti 16 GB: \~24 s per edit at 1.5K, \~40 s for a native 2K edit.
* **Agents:** built-in MCP server (`generate_image`, `edit_image`, `inpaint_image`, `outpaint_image`, `upscale_image`…) and a local HTTP API: `npx -y kokon-mcp`, also in the official MCP Registry.
* **Not yet:** Linux; code signing (SmartScreen warns on first run); reusing model files you already have is a manual workaround for now ([\#14](https://github.com/GabrieleLozupone/kokon/issues/14)).

Repo: [https://github.com/GabrieleLozupone/kokon](https://github.com/GabrieleLozupone/kokon) · Site: [https://kokon.app](https://kokon.app)

Code is AGPL-3.0; the Qwen weights are under the Qwen Research License (non-commercial).

document fetched 2026-10-15 00:45 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5wyii/gptimagestyle_photo_editing_on_your_own_gpu_12_gb