---
name: image-retouching
description: Retouches images on your own GPU with commercially licensed models — background removal, upscaling, removing a logo or object, masked inpainting, stills generated locally. Use for a transparent PNG, upscale, removing an overlay from your own image, inpainting with a mask, ComfyUI or VRAM limits, or a model like FLUX, Z-Image, RMBG, BiRefNet or LaMa.
---

# Image retouching

## Rules for any model

- **License of the weights first.** The most-cited model for a task is often
  non-commercial. Check the model card before it touches company work; the
  models this skill uses, and the popular ones to avoid, are in
  [references/models.md](references/models.md).
- **Removing is not inpainting.** To take something out (a logo, a sign, a
  watermark), use a texture-fill model such as LaMa: it rebuilds the
  surroundings and invents nothing. Diffusion inpainting asked for "clear sky"
  over trees invents objects. Use prompted inpainting only to put something
  new in.
- **One GPU job at a time** on an 8 GB card: a diffusion model in fp8 takes
  about 6 GB, so a TTS or music model beside it overflows.
- **Video upscaling** does not pay off on 8 GB (hours per minute of video, and
  it invents detail). Use it only to rescue third-party low-resolution footage.

## With the bundled ComfyUI scripts

ComfyUI is an HTTP server; the UI is one client. `scripts/comfy-run.py` starts
the server through `scripts/comfy-serve.sh` when none is running, posts an
API-format graph from `workflows/`, waits on `/history`, downloads the result
and stops what it started. From this skill's directory:

```bash
export COMFY_HOME=~/.local/opt/comfyui        # holds ComfyUI/ (clone) and .venv/
python3 scripts/comfy-run.py background photo.png -o cutout.png
python3 scripts/comfy-run.py upscale    photo.png -o 2x.png --factor 2
python3 scripts/comfy-run.py remove     frame.png -o clean.png --rect 1500,60,360,120
python3 scripts/comfy-run.py inpaint    photo.png "what to put there" -o out.png --mask m.png
python3 scripts/comfy-run.py generate   "description" -o still.png --size 1344x768 --seed 7
```

Pass `--keep` on the first of several jobs; the first model load is slow.

- The JSON the UI saves (`nodes`, `links`, subgraphs) is not what `/prompt`
  accepts. The workflows here are API format (`{"id": {"class_type",
  "inputs"}}`); convert a UI template with "Export (API)". Recent official
  templates hide nodes inside `definitions.subgraphs`.
- Fit 8 GB: start the server with `--fp8_e4m3fn-unet` (bf16 converts to fp8 on
  load) and reuse a cached `torch` wheel through `uv`.

Install and model files: [references/install.md](references/install.md).