Chibi Character Conversion · skill pública Open Skill

Chibi Character Conversion

Chibi Character Conversion é uma Open Skill pública no Digen. Abra para ver o que este fluxo cria.

@wwws

Anime

0 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/fpyeeg3x.jpg

# Chibi Character Conversion

Turn one uploaded photo into a single cute chibi illustration: same person, same recognizable features, redrawn in a pastel super-deformed anime style. There is no structured project state — the uploaded asset plus the conversation are the only inputs.

## When to use

- The user uploads a photo (selfie, portrait, cosplay, character art) and asks for a "chibi", "SD", "cute cartoon", "Q-version", "anime version", or "sticker" version of it.
- The user uploads a photo and asks to redraw it in a cute pastel anime style, optionally changing the outfit, pose, or background theme.
- The user asks for a chibi avatar / profile picture based on an uploaded photo.

## Inputs

| Input | Required | How it is used |
|---|---|---|
| Reference image | Yes | Passed as `input_urls` to `generate_image` with `model='i2i.hd.lite'`; the single source of identity (face, hair, colors, accessories). |
| Text | Optional | Theme, outfit, pose, mood, or background direction. Translated into the outfit / pose / background slots of the prompt, never into identity changes. |

If the user only sends text and no image, ask once for the photo. If they insist on text-only, generate with `model='t2i'` and tell them it is a new character, not a likeness conversion.

## Steps

1. **Check the inputs.** Confirm an image asset is present. Note the text request (theme / outfit / pose / background). If the request contradicts the fixed style (e.g. realistic photo, 3D, gore), state the limitation before generating.

2. **Read the identity from the image.** Call `recognize_image` with the uploaded asset as `image_url` and a prompt asking for identity-defining details: face shape, eye shape and color, eyebrows, hairstyle (length, bangs, parting), hair color, glasses, moles, signature accessories, outfit items and their colors. Use the returned list as the identity lock. If the call fails, use only details that are clearly visible or ask the user briefly — never invent features.

3. **Read the references before generating.**
   - `read_skill_file` → `references/style-and-prompt.md` (style, subject, layout, must-keep, must-avoid, prompt skeleton).
   - `read_skill_file` → `references/step-details.md` (checklist, orientation decision, prompt assembly order, self-review and retry rules).

4. **Assemble the prompt** by filling the template in `references/style-and-prompt.md`: identity lock first, then outfit and pose (from the user's text or the source photo), then background, then the fixed style tokens. Write the whole prompt in English.

5. **Generate the image.** Call `generate_image` with:
   - `prompt` = the assembled English prompt,
   - `input_urls` = the uploaded reference asset,
   - `model='i2i.hd.lite'`,
   - `orientation='portrait'` for a full-body standing character (default), `'square'` for an avatar / head-and-shoulders request, `'landscape'` only if the user asks for a wide scene,
   - `name` = short title, e.g. `Chibi <subject> <theme>`,
   - `entity_type='characters'`.

6. **Self-review the result** against the must-keep list in `references/style-and-prompt.md`. If identity drifted (hair, eye color, glasses, signature accessory lost), or proportions look realistic instead of chibi, retry at most twice with a stronger identity-lock sentence and fewer background motifs. If it still fails, report what could not be matched instead of claiming success.

7. **Reply.** Describe the result in one or two natural sentences (what was kept from the photo, what changed). Do not paste long URLs. Offer 2–3 concrete next options via `set_guidance` in the same turn, e.g. square avatar version, a different outfit theme, a different pastel background.

## Notes

- This is a single-image conversion Skill. Deliver exactly one illustration; do not promise a before/after collage or photo comparison — the tool cannot composite images.
- Keep the subject's identity, ethnicity, skin tone, age bracket, and gender presentation as they appear in the source