AI image editing inside GIMP with gimp-comfy-tools

2026-09-02

What happened

I wanted to do AI image generation and inpainting without leaving GIMP, so I wired up gimp-comfy-tools — a GIMP 3.0 plugin suite that talks to a local ComfyUI server over its REST + WebSocket API. It opens a "ComfyUI Studio" panel inside GIMP where you type a prompt, pick a checkpoint, and the result comes back as a new layer with all the generation parameters stored on the layer as metadata.

The setup turned out to be more involved than dropping a folder into a plugin directory, so here's the whole path in case it helps someone else.

The stack: GIMP 3.2.4 (AppImage), a local ComfyUI checkout with a fresh Python venv, PyTorch CUDA build for my RTX 4070 Super, and the plugin suite copied into GIMP's user plug-ins directory.

Three gotchas that ate the most time:

1. The right plug-ins directory. GIMP 3.2 keeps its config in ~/.config/GIMP/3.2/ (not 3.0), and the plugin folder must sit there as three siblings — modules/, gimp-generate-pers/, and gimp-metadata-viewer/. Drop it in the wrong profile dir and nothing shows up, no error.

2. The Python shebang. The plugin's #!/usr/bin/env python3 resolved to my pyenv Python, which had no gi bindings. Installing gir1.2-gimp-3.0 system-wide and pointing the shebang at /usr/bin/python3 fixed the "no panel at all" symptom.

3. The beige-gradient bug. The first several generations came back as a flat beige gradient. The culprit: the plugin's persisted settings had saved cfg: 0.0 and denoise: 0.0 (out-of-range zeros that slipped past the UI spinners). With zero CFG guidance the model just produces a flat field. Clamping cfg/denoise/seed to sane values at read time fixed it, and suddenly real images came out.

What finally worked: a RealisticVision v5 (SD 1.5) checkpoint dropped into models/checkpoints/, two example workflows registered, and the settings clamped so degenerate zeros can't propagate. A test generation did 25 steps at ~4.9 it/s on the 4070 Super — a full image in under 9 seconds including model load.

Files touched

• ComfyUI checkout: ~/sandbox/thirdparty/ComfyUI (venv + PyTorch CUDA + checkpoint)
• GIMP plug-ins: ~/.config/GIMP/3.2/plug-ins/{modules,gimp-generate-pers,gimp-metadata-viewer}
• GIMP comfy data: ~/.config/GIMP/3.2/comfy/Workflows/ (workflow registry + example workflows)
• GIMP comfy data: ~/.config/GIMP/3.2/comfy/data/last_inputs_persistent.json (persisted settings)

Tweet draft

Piped Stable Diffusion into GIMP with gimp-comfy-tools + a local ComfyUI. Text-to-image and inpainting right inside the editor, results come back as layers tagged with their generation params. The fun parts: the plug-ins dir that wasn't where I thought, the missing GI bindings, and chasing a flat beige gradient back to a cfg: 0.0 that leaked into persisted settings. Now it does 25 steps in ~5s on a 4070 Super [link]

commit: 7041849 · screenshot: no screenshot