Witch Experience · Open Skill 公開スキル

Witch Experience

Witch Experience は Digen 上の公開 Open Skill です。開くと、このワークフローが何を作るか分かります。

@Promise

Manhua

0 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/3949831/2026-10-14/h2kwdpr2.jpg

# Witch Experience

Turn the **main subject** in the user’s reference photo into a **witch flying on a broomstick** against an open sky, wearing recognizable **witch costume** (pointed hat, cloak/dress, optional gloves/boots). Preserve the subject’s **face, body type, hair, and skin tone** from the upload; change outfit and scene to match the fantasy.

## When to use

- User uploads a photo (portrait, half-body, or full-body) and wants a **witch on a broom** result.
- User mentions witch cosplay, Halloween witch, flying broom, sky background, or **女巫体验**.
- User wants another take, higher quality, or a different I2I tier on the **same** reference image.

## Inputs

- **Required:** at least one reference image from the user (attachment or canvas asset). Use its URL / `asset_id` / `s3://` uri in `input_urls`.
- **Optional:** sky mood (sunset, storm, full moon), outfit style, colors, cartoon vs photoreal — follow user wording.

## References

| File | When to read |
|------|----------------|
| `references/witch_attire.md` | Before composing the I2I prompt (step 4): witch **style presets**, accessories, broom, sky, and color palettes |

Do not load this file on unrelated turns. Use `read_skill_file(witch-experience, references/witch_attire.md)`.

## Steps

1. **Confirm reference**  
   If no image is available, ask the user to upload one. Do not call `generate_image` without `input_urls`.

2. **Analyze subject (recommended)**  
   Call `recognize_image` on the reference with a prompt like:  
   *Describe the main person only: apparent age, gender presentation, face shape, hair, skin tone, pose, and clothing colors. Ignore background.*  
   Use this to keep likeness in the I2I prompt.

3. **Pick I2I channel (required — only these three)**  

   | Situation | `model` |
   |-----------|---------|
   | Default first generation, fast preview | `i2i.hd.lite` |
   | User asks for best quality, sharper detail, or HD | `i2i.hd` |
   | User names GPT tier, or lite/HD failed twice on likeness | `i2i.hd.gpt` |

   Do **not** use bare `i2i`, `t2i*`, or physical model names. If unsure, call `list_models` with `model_type='image'` only to confirm these channels exist — do not override the table without user request.

4. **Outfit & scene (reference)**  
   Call `read_skill_file(witch-experience, references/witch_attire.md)`. Pick one **style preset** (or match the user’s words), plus broom variant and sky row. If the user gave no style preference, use **Default bundle** at the bottom of that file.

5. **Generate**  
   Call `generate_image` with:

   - `input_urls`: `[reference]` (first image = subject to preserve)
   - `model`: chosen channel from the table
   - `orientation`: `portrait` if the upload is vertical or head-and-shoulders; `landscape` if clearly wide; `square` only when the user wants 1:1
   - `name`: short title, e.g. `Witch Experience` or `Witch Flight`
   - `entity_type`: `characters`
   - `prompt`: English; merge likeness + chosen blocks from `witch_attire.md` + checklist at end of that file

   **Prompt skeleton (fill from steps 2 and 4):**

   ```text
   Image-to-image: keep the same person as the reference — [likeness from recognize_image].
   Transform them into a witch riding a broomstick flying high in the sky, dynamic flying pose, legs straddling the broom, wind in hair and cloak.
   [Witch costume: paste 2–4 sentences from chosen style preset + any accessories from reference tables.]
   [Broom: one line from Broomstick variants.]
   [Background: one row from Sky & lighting pairings.]
   Cinematic lighting, high detail, coherent anatomy, single subject focus, family-friendly.
   No text, no watermark, no extra people, no gore.
   ```

6. **Reply**  
   Summarize what changed in plain language (mention witch **style** if not classic). Do not paste long asset URLs. If the result lost likeness, retry once with the **next tier up** in the table (lite → hd → gpt) before asking the user.

7. **Optional follow-ups**  
   After s