CharaKit
Edit visual novel character illustrations in Codex. This plugin provides expression variants, clothing edits, and accessory / prop edits with versioned candidates, human review, and selected PNG exports.
Development version 0.1.28 · Repository · 中文介绍
Codex generates artwork through its available image editing tool. The bundled Python helper preserves raw files and manages comparisons, review records, revisions, and packaging.
Skills
| Skill | Supported edits |
|---|---|
$charakit-expressions | Twelve expression presets and static mouth-open / mouth-closed variants |
$charakit-outfits | recolor, replace, outfit_replace, and prop_edit with add, remove, or replace actions |
The modules use independent projects. To create expressions after changing an outfit, manually use the resulting PNG as a new expression source. Pose editing remains planned.
Setup
You need Codex plugin support, an available image editing tool, and Python 3.11+. Install the helper dependency from this plugin directory:
python -m pip install -r requirements.txt
Keep the complete plugin directory, including lib/, references/, and both Skills. Use the Python environment with Pillow installed. For marketplace installation, follow the repository instructions or the official packaging guide.
Quick start
Attach a source illustration in Codex and ask:
Use $charakit-expressions to create happy and worried expressions,
each with mouth closed and naturally open for speaking.
Keep the character's facial design, hair, clothing, pose, and original style.
Or edit clothing or an object:
Use $charakit-outfits to replace the jacket with a short navy blazer.
Keep the shirt, skirt, shoes, accessories, face, hair, expression, and pose.
Use $charakit-outfits to add thin-frame glasses.
Keep the facial features, expression, hair, clothing, and pose.
For a full outfit, name all included garments, socks, and shoes. Unlisted components remain protected. Optional reference images can guide explicitly named attributes in either module; state what each reference is for.
Responses and preview UI follow your current or preferred language. Machine-readable IDs, filenames, and CLI arguments remain stable English identifiers. Artwork belongs in your workspace, normally art-output/<character-key>/, outside the installed plugin.
Review and delivery
- Each requested result gets up to three attempts, including the initial image, unless you set a smaller limit. Internal failures trigger redraws within that allowance; full passage stops generation early.
- Internal passage requires 80/100 or higher, all five dimensions at least 4/5, and all applicable domain, fidelity, and file checks. Scores are Codex's visual observations, not automated image measurements.
- After passage or exhaustion, the local HTML review page shows every readable candidate in the batch. Full images, optional face / target comparisons, and synchronized source viewing support inspection. Recommendations do not preselect a candidate.
- Reply “Use A”, “Revise B: …”, or “Reject all” in chat, or copy feedback from the page. Corrections edit the chosen version, preserve successful changes and recheck still-protected regions against the original. Revision previews default to before/after comparison with task-based local views and a switchable source baseline. Failed candidates remain open to feedback and revision.
- Actual human feedback is required for the final review report. Only accepted, eligible candidates become final full PNGs, a comparison preview, and a requested ZIP. Final reports also include recorded face / target inspection images separately from the full artwork. Earlier reports are internal drafts.
Positive feedback cannot bypass failed delivery checks. If every attempt fails, the workflow explains the remaining issues and retains the candidates for human review rather than exporting them as finished art.
Helper entrypoints
Codex uses the helper during the workflow; users can work through chat. To inspect the available commands, run from this plugin directory:
python skills/charakit-expressions/scripts/studio.py --help
python skills/charakit-outfits/scripts/studio.py --help
Outside this directory, use the entrypoint's absolute installed path. Both modules share lib/studio_core.py. Commands manage source snapshots, raw candidates, comparisons, assessment records, actual feedback, revisions, and exports. Work sequentially within one project.
Guides
| Topic | Documentation |
|---|---|
| Expression workflow | Expressions Skill · Presets · Mouth states |
| Clothing and object workflow | Outfits Skill |
| Clothing modes | Recoloring · Single replacement · Full outfits |
| Accessories and props | Prop edits |
| Reference roles and protection | Scoped references · Editing boundaries |
| Candidate review and revisions | Human review and redraws · Quality policy |
| Final reports and exports | Delivery reports |
The guides include command details and legacy compatibility rules. See GitHub Releases for release history.
Limits
Inputs are static PNG/JPEG images up to 50 MiB and 40 million pixels. Final assets must be PNGs matching the source canvas and passing applicable transparency and integrity checks. Canvas mismatch may be shown for review, but blocks formal acceptance and export.
The helper preserves generated image bytes; it does not generate, resize, repaint, composite, or change alpha. Image editing can still alter protected details, so visual inspection remains necessary. CharaKit supplies no image model, external image API, MCP service, animation, or lip synchronization.