Halloween Dark Transformation · compétence Open Skill

Halloween Dark Transformation

Halloween Dark Transformation est une Open Skill publique sur Digen. Ouvrez-la pour voir ce que ce workflow crée.

@JackRhino

Roleplay

1 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-20/utjqtbrd.jpg

# Halloween Dark Transformation

Turn one uploaded photo into a Halloween dark-transformation portrait: the same person, the same place, the same pose and camera distance, restyled with a witch / vampire / witch-vampire look under low-key Halloween lighting. Single deliverable: one image.

There is no structured project state. Everything the skill needs lives in the user's photo, the user's text, and the generated canvas assets.

## When to use

- The user uploads a selfie or portrait and asks for a Halloween version, a dark Halloween look, a witch or vampire costume, Halloween makeup, or a "dark transformation" of the photo.
- The user asks to revise an already generated Halloween portrait (stronger makeup, different costume, darker light, keep the face more faithful, etc.).
- The user uploads a photo plus text such as "make me a vampire", "more witch, less blood", "halloween party poster vibe".

Do not use for: video requests, group photos of several unrelated people, non-Halloween cosplay, or pure text-to-image without any photo (see Notes for the fallback rule).

## Inputs

| Input | Required | How it is used |
|---|---|---|
| Reference image (selfie / portrait) | Yes | Passed as `input_urls` to `generate_image`; also the source of identity, pose, framing and location that must be preserved. |
| Text (user request) | Optional | Selects the look (witch / vampire / hybrid), intensity, costume details, props and mood. Absent text means the default hybrid witch-vampire look at standard intensity. |

## Steps

1. **Intake.** Identify the reference image and the user's text. If no image is attached, ask for one upload (this skill is image-to-image; identity comes from the photo). If several images are attached, ask which one is the subject. Full intake checklist: `references/step-details.md` (Step A).
2. **Read the visual contract.** Call `read_skill_file` with `references/style-and-prompt.md` and `references/transformation-looks.md` before writing anything. Do not improvise style rules from memory.
3. **Map intent to a look.** Use the keyword-to-look table in `references/transformation-looks.md` to pick witch / vampire / hybrid and an intensity level (subtle / standard / full). State the pick in one short line for the user.
4. **Optional photo reading.** When the photo is ambiguous (small subject, heavy filters, unusual framing, multiple people in frame), call `recognize_image` on the reference to record subject description, pose, camera distance, background and existing light. Use its notes only as input for identity preservation.
5. **Write the English prompt.** Assemble the prompt with the template and fragment libraries in `references/style-and-prompt.md`, filling the look-specific lines from `references/transformation-looks.md`. Follow the assembly order and length rules in `references/step-details.md` (Step C). Never ship a bare one-line prompt.
6. **Generate.** Call:

   `generate_image(prompt=<assembled English prompt>, input_urls=[<reference image>], model='i2i.hd.lite', orientation=<mirror the source aspect ratio>, name='Halloween Dark Portrait - <look>', entity_type='characters')`

   - `model` is always the image-to-image channel `i2i.hd.lite` for this skill.
   - `orientation`: `portrait` unless the reference is clearly landscape (`landscape`) or square (`square`); never crop the person's framing differently from the source.
   - Leave `resolution` unset by default; pass `resolution='1080P'` only when the user asks for high detail or print use.
   - One image per call (`batch_size` stays 1). For several looks, call the tool once per look.
7. **Quality check.** Run the checklist in `references/step-details.md` (Step E): identity still recognizable, pose and framing unchanged, Halloween set dressing present, low-key light with readable highlights, no unwanted text/watermark, no NSFW content. If a check fails, revise the prompt and regenerate following the branch rules in Step F instead of shipping a failed image.
8.