Mark Pen Graffiti Chibi (I2I) · Open Skill 공개 스킬

Mark Pen Graffiti Chibi (I2I)

Mark Pen Graffiti Chibi (I2I)

@ExploreFly

Other

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3656543/2026-10-14/s3bs6dac.jpg

# Mark Pen Graffiti Chibi (Image-to-Image)

Turn the user's uploaded photo into a **chibi (super-deformed) hand-drawn mark-pen graffiti** illustration while keeping recognizable hair, outfit, and key accessories from the source.

Structured project state is not used; progress is the conversation plus canvas assets.

## When to use

- User asks for **Mark Pen Graffiti**, **marker doodle**, **chibi graffiti**, or style like **Qwen Image 2.1 Mark Pen Graffiti**.
- User uploads a reference photo and wants a **stylized cartoon / SD character** version, not a realistic edit.
- User wants to **retry**, **tweak doodle intensity**, or **switch I2I quality tier** on the same photo.

## Style reference (read before generating)

Full taxonomy, prompt blocks, and quality checklist live in:

| File | When to read |
|------|----------------|
| `references/mark-pen-graffiti-style.md` | Before every `generate_image` call (prompt assembly) |

Do **not** guess prompt wording from memory; `read_skill_file(name='mark-pen-graffiti-chibi', path='references/mark-pen-graffiti-style.md')` first.

## Workflow

1. **Input** — Confirm the user attached at least one image. Use that asset's URL / `asset_id` / `s3://` URI as the sole reference in `input_urls` (first image = identity + outfit source). If multiple images, prefer the clearest front-facing subject unless the user names another.
2. **Analyze (recommended)** — `recognize_image` on the reference with a short prompt: list hair color/style, skin tone, clothing layers, tie/scarf/jewelry, pose, and any props. Use this only to fill the **Subject preservation** slots in the style reference; do not copy photographic lighting verbatim.
3. **Prompt** — Merge the **Master positive prompt** from the reference file with subject-specific lines (hair, outfit, accessories). Keep the full generation prompt in **English**.
4. **Orientation** — Match the user's intent: default **`square`** for a centered chibi character on paper; use **`portrait`** if they want a taller full-body doodle; use **`landscape`** only if they explicitly ask for horizontal layout. Omit `resolution` unless the user requests higher output (then prefer `1080P` on `.hd*` channels).
5. **Generate** — Call `generate_image` with `input_urls=[reference]`, assembled `prompt`, chosen `model` channel (below), `name` like `Mark pen graffiti chibi`, `entity_type='characters'`.
6. **Reply** — Summarize what was preserved and the doodle style in one or two sentences. Do not paste long asset URLs.

## I2I model channels (required)

This skill uses **image-to-image only**. Pick one channel per request:

| Channel | When to use |
|---------|-------------|
| `i2i.hd.lite` | **Default** — fast preview, first attempt, or user did not ask for max quality. |
| `i2i.hd` | User wants **sharper marker texture**, cleaner edges, or says "HD / high quality". |
| `i2i.hd.gpt` | User wants **best quality**, "GPT", or maximum fidelity to the graffiti chibi look after `.hd` is still not enough. |

Rules:

- Always pass **`model`** explicitly as one of: `i2i.hd.lite`, `i2i.hd`, `i2i.hd.gpt`.
- Never pass physical model names (e.g. `qwen-image`); channels only.
- All three `.hd*` channels **do not support NSFW** — refuse adult/explicit requests politely without calling `generate_image`.
- If a channel fails, report the error; optionally call `list_models` with `model_type='image'` once to confirm I2I channels, then retry with the next tier if the user agrees.

Example call shape:

```text
generate_image(
  prompt=<assembled English prompt>,
  input_urls=[<user reference>],
  model='i2i.hd.lite',
  orientation='square',
  name='Mark pen graffiti chibi',
  entity_type='characters'
)
```

## Iteration

- **More scribbly / messier lines** — strengthen "messy marker outline", "visible hatch strokes", "ink bleed" phrases from the reference; stay on `i2i.hd.lite` or bump to `i2i.hd`.
- **Cleaner chibi** — reduce "messy" modifiers slightly; prefer `i2i.hd` or `i2i.hd.gpt`.
- **More