# Inpaint Images with RunComfy

Precise image edits are hard when only one region should change. This skill routes Claude, Codex, or Claude Code to RunComfy inpainting models with mask-aware prompts and safe CLI usage.

## Install

```bash
npx skillstore add agentspace-so/image-inpainting
```

## Metadata

- Status: approved
- Slug: agentspace-so-image-inpainting
- Skillstore revision: r1
- Version status: missing
- Tree hash: 09040e6462680530ee475841864300629c543731f3cd726c2de2d435b2a8dfd9
- Author: agentspace-so
- GitHub username: agentspace-so
- License: MIT
- Repository: https://github.com/agentspace-so/runcomfy-agent-skills/tree/main/image-inpainting/
- Ref: 1ffa7643651792ccb4bd3b15d924d2c97edff755
- Supported tools: Claude, Codex, Claude Code
- Audit status: complete
- Agent install advisory: allowed
- Manual install advisory: allowed
- Artifact signature: available
- Audit attestation: unavailable
- Human verification: not\_verified
- Risk factors: external\_commands, network, filesystem
- Quality score: 70
- Public page: https://skillstore.pages.dev/skills/agentspace-so-image-inpainting
- Manifest: https://skillstore.pages.dev/api/skills/agentspace-so-image-inpainting/manifest

## Capabilities

- Runs RunComfy inpainting jobs with source image, mask image, prompt, strength, and control scale inputs.
- Recommends Z-Image Turbo Inpainting when a binary mask is available.
- Routes no-mask edits to description-based image edit models such as Nano Banana 2 Edit, GPT Image 2 Edit, or FLUX Kontext Pro.
- Provides parameter guidance for retouching, object replacement, and full region replacement.
- Explains expected CLI exit codes and troubleshooting categories.

## Use Cases

- Clean product photos: Replace a masked background or remove a small defect while preserving the product shape and lighting.
- Repair image regions: Fill a damaged or empty area using a mask and a prompt that describes the desired replacement.
- Localize visual edits: Use description-based editing when a mask is not available but the target region can be named clearly.

## Prompt Templates

### Remove a masked object

```
Inpaint this image using the provided mask. Remove the masked object and preserve the background texture, lighting, and perspective.
```

### Replace a masked background

```
Use the mask to replace only the background with smooth light gray studio paper. Preserve the product edges, shadows, and color.
```

### Choose the best edit route

```
I have an image and this edit goal, but I am not sure if my mask is good enough. Choose the best RunComfy inpainting or edit route and explain the needed inputs.
```

### Tune an inpainting job

```
Run an inpainting job for this masked region. Use a lower strength for subtle repair, set control scale for edge preservation, and include a reproducible seed.
```

## Limitations

- Requires the runcomfy CLI, an authenticated RunComfy account, and network access.
- Mask-driven inpainting requires a separate image mask URL.
- The skill does not create masks inside the workflow.
- Generated quality depends on the source image, mask quality, model choice, and prompt clarity.

## Best Practices

- Use a clean grayscale mask where white marks the region to change and black marks the region to preserve.
- Name the surrounding details that must stay unchanged in every prompt.
- Use only images and marks you own or have permission to edit.

## Anti Patterns

- Do not use a vague prompt that fails to describe what should remain unchanged.
- Do not reuse a mask from a different source image.
- Do not submit private or regulated images unless remote processing is approved.

## Security Audit

- Audited at: 2026-07-09T05:40:41.705\+00:00
- Summary: Most static findings are false positives caused by Markdown backticks, documentation links, placeholder URLs, and explanatory text. The real security surface is limited to intentional runcomfy CLI execution and outbound uploads to RunComfy, plus a semantic misuse risk from explicit watermark-removal guidance.

## Stats

- Views: 59
- Downloads: 9
- Favorites: 0
- Popularity score: 0