im2-image-skills

KO | EN

gitlite — search

#deep-learning #graphql #computer-vision #self-hosted #generative-ai #command-line #dotnet #mcp #tailwindcss #automation

im2-image-skills

(★ 9)

Open-source Codex skills for clean IM2 image generation, material realism, negative-prompt hygiene, and grounded scene mass.

File Explorer

ZIP Download

- im2-clean-image-one-page.png
- im2-clean-image-one-page.svg
- - SKILL.md
.gitignore
LICENSE
README.md

# Use via CDN

jsDelivr

jsDelivr serves any public GitHub repository as a CDN with zero setup. Pick a version and a file to get a ready-to-paste link and snippet.

Version

File

Link

Example

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

A compact, reusable Codex skill collection for clean IM2 / GPT Image 2 image generation and editing.

![IM2 Clean Image one-page workflow](https://raw.githubusercontent.com/q2522879285-source/im2-image-skills/main/assets/im2-clean-image-one-page.png)

## Included skill

### `im2-clean-image`

Adds four core control layers:

- physically distinct material and light behavior;
- information-density hierarchy with controlled information loss;
- clean rendering and anti-artifact controls;
- targeted negative-prompt hygiene.

Conditional modules add structural mass for heavy scenes, evidence-preserving photo repair, and real-scene light paths, light ratio, air scattering, atmospheric perspective, conditional volumetric light, and scale-aware texture.

It also includes a conservative evidence-preserving branch for documentary, event, team, product, food, and document-photo restoration.

## What the real-scene update adds

The skill now compiles physically observable scene logic before mood language:

1. separate scene invariants from aesthetic freedom;
2. trace `source → medium → receiver`, with camera/viewing geometry controlling what scattering and reflection are visible;
3. distinguish atmospheric perspective from visible volumetric light;
4. build contrast without crushed blacks or flat fill;
5. reveal texture only when camera scale, light, and narrative importance support it;
6. inspect the actual output, freeze passing controls, and change one main variable per retry.

## Install

For a fresh install, copy the skill folder into your Codex skills directory:

```powershell
Copy-Item -Recurse .\skills\im2-clean-image "$HOME\.codex\skills\"
```

To update an existing install without creating a nested directory:

```powershell
$target = "$HOME\.codex\skills\im2-clean-image"
New-Item -ItemType Directory -Force $target | Out-Null
Copy-Item -Recurse -Force .\skills\im2-clean-image\* $target
```

Restart or reload Codex after installation. Then invoke it explicitly:

```text
Use $im2-clean-image to rewrite this image prompt with clean material realism.
```

Codex may also invoke it implicitly for matching IM2 image-generation and image-editing requests.

Verify the installation by confirming that Codex lists `im2-clean-image`, then try:

```text
Use $im2-clean-image to classify this source-image request as EDIT, REBUILD, or MIXED and compile the image transaction.
```

The skill is self-contained. If the optional companion `$constraint-input-router` is installed, it can supply the edit/rebuild classification; otherwise `im2-clean-image` uses its bundled minimal classification fallback.

## Main files

- `SKILL.md` — core workflow and routing
- `references/clean-prompt-recipes.md` — task-specific prompt recipes
- `references/negative-prompt-hygiene.md` — compact avoid-block discipline
- `references/evidence-preserving-photo-restoration.md` — factual photo-repair constraints
- `references/heavy-scene-structure-and-mass.md` — architecture and hard-surface mass, acceptance, and retry loop
- `references/real-scene-light-atmosphere.md` — source/medium/receiver light paths, scattering, volumetrics, light ratio, material hierarchy, and retry checks

## Design principles

- Preserve visible facts before styling.
- Describe material physics instead of stacking generic quality spells.
- Keep large surfaces quiet and detail scale-dependent.
- Treat air and volumetric light as physical media, not decoration.
- Treat exact preservation as an acceptance target, not a guarantee of generative control.
- Inspect actual outputs before declaring success.
- Freeze passing controls and change one main variable per retry.
- Keep platform settings outside the content prompt unless an adapter requires prompt syntax.

## Scope

This repository provides prompt-compilation and review instructions. It does not include a model, API client, platform account, or generated assets. Exact text, logos, registered layouts, or measurable geometry may require masks, compositing, registration, comparison, or 3D reconstruction.

## License

MIT. See [LICENSE](LICENSE).

im2-image-skills

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

im2-image-skills

Owner Repos

File Explorer

# Use via CDN

Was this content helpful?

Recent Feedback

Download README