---

name: gimp-ml description: "Route GIMP-ML computer-vision plug-in, GIMP layer, local FastAPI image-generation, and legacy-host troubleshooting tasks with explicit model, runtime, credential, and verification boundaries." disable-model-invocation: true metadata: disco-role: operating license: MIT

GIMP-ML

Use this repo skill when a task names GIMP-ML or asks for GIMP-integrated computer vision such as deblurring, denoising, dehazing, enhancement, depth, segmentation, face parsing, matting, inpainting, super-resolution, frame interpolation, K-means, palette operations, or text-to-image/edit/extend/outpaint. This is a router for the repository's two distinct surfaces: historical GIMP 2 Python-Fu plug-ins and a newer local Python 3 HTTP service.

First classify the task

1. Start with setup-and-host for installation, host compatibility, plugin discovery, weights, permissions, refresh safety, GIMP menus, or choosing between legacy and service paths.
2. Use classical-image-ops for invert, K-means clustering, palette behavior, drawable/array checks, and CPU-safe image primitives.
3. Use vision-filters for deblur, dehaze, denoise, enhancement, monocular depth, semantic segmentation, face parsing, super-resolution, or frame interpolation.
4. Use guided-editing for inpainting, trimap matting, face-generation layer preparation, or sparse color masks. Route ordinary segmentation or restoration back to vision-filters.
5. Use text-generation-service for the local FastAPI protocol, GIMP 2 bridge payloads, text-to-image, text-edit, text-extend, outpaint, or OpenAI/provider failures.

Read architecture when a request spans routes, and troubleshooting before changing a host, weights tree, service process, or configuration. Run the shared environment diagnostic for a read-only import, GIMP/Python-2 executable, or no-allocation CUDA report.

Verification boundary

The generated operating context is based on public docs and source evidence at one repository snapshot. A Python 3 service-core verification established common imports, package consistency, and FastAPI route declarations. The verification host had no GIMP or Python 2.7, no model weights, and no provider credentials. CUDA was visible but a tiny allocation was blocked by the host's CUDA memory state. Therefore:

• A static registration scan or input-validator pass is not GIMP execution.
• A present checkpoint path is not compatible weights or successful inference.
• CUDA visibility is not proof that a filter can run on the target image.
• A local /status response is not model loading or provider success.
• No skill instruction downloads weights, calls OpenAI, or runs a destructive updater automatically.

Keep those labels in any downstream report. Read repo provenance before treating this graph as current for another checkout.

Safe operating sequence

1. Identify the requested surface and required inputs (selected layer, aligned mask/trimap, second frame, output directory, or HTTP payload).
2. Run the nearest bundled read-only validator or asset/layout check before loading a model or mutating GIMP state.
3. Check the host/backend gate: legacy workflows need a compatible GIMP 2.10 Python-Fu/Python 2 host; model workflows also need external weights; text workflows need an installed service launcher and separately authorized provider access.
4. Preserve source layers and input files. Report whether the result is generated, static-only, blocked, or unverified rather than inferring success.

There is no single installable Python distribution for the legacy plugin collection. For a read-only local diagnostic, run the bundled helper from the root of this generated skill:

python scripts/check_environment.py --help

Install or launch the legacy GIMP application, model assets, or service only through the target deployment's separately reviewed public procedure.