Leather Outfit Transform · publiczny skill Open Skill

Leather Outfit Transform

Leather Outfit Transform to publiczny Open Skill w Digen. Otwórz go, by zobaczyć, co tworzy ten workflow.

@wwws

TryOn

1 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/kgu8uxb6.jpg

# Leather Outfit Transform

Rebuilds the outfit in an uploaded photo as a glossy hard-leather version of itself — candy-pastel patent leather, chrome hardware, wet-look highlights — while the person's identity, hair, pose, framing and background stay unchanged. There is no structured project state: the reference image, the finished edit and the conversation are the whole record.

## When to use

- The user uploads a photo of a person and wants the clothing turned into leather / latex / PVC / patent-leather fashion.
- The user asks for a "leather version", "hard leather look", "costume change to leather", "same person, new leather outfit".
- The user already got a result and now wants a variation (different leather colour, more / less hardware, higher gloss, another pose kept as-is).
- Not for: creating a brand-new character, changing the person's face or scene, or generating an outfit on a mannequin / product shot with no person in the reference.

## Steps

1. **Read the reference image first.** Before calling any tool, note down in one short list:
   - identity anchors: face shape, eyes, brows, nose, mouth, hairstyle, hair colour, fringe, makeup mood, skin tone;
   - the outfit's colour story and design cues (e.g. "pink + baby-blue side panels, white belt, one-shoulder leotard");
   - pose, camera angle, framing (full body / three-quarter / close-up), background and light direction.
   If the photo is not a person, is a flat product shot, or the outfit is unreadable, say so and ask for one clear photo instead of guessing.

2. **Load the visual rules.** Call `read_skill_file(name='leather-outfit-transform', path='references/style-and-prompt.md')` and follow it: it holds the Style, Subject/content, Layout, Must-keep, Must-avoid blocks plus the copy-ready English prompt templates. Do not invent the look from memory — the file is the source of truth.

3. **Load the checks when needed.** If the input is ambiguous (multiple people, low resolution, cropped body, non-person subject) or a previous attempt failed the verification in step 6, call `read_skill_file(name='leather-outfit-transform', path='references/step-details.md')` for the intake rules and the retry ladder.

4. **Write the prompt in English.** Take Template A (single transformed portrait) from `references/style-and-prompt.md` and fill every placeholder with the values from step 1: the leather garment build, the preserved colours, the hardware, the gloss level, the scene description. Append the negative list from the same file. Only switch to Template B (before/after comparison card) when the user explicitly asks to see "before and after", a comparison, or a poster-style card.

5. **Generate with the image-to-image channel.** Call `generate_image` with:
   - `prompt`: the filled English prompt from step 4;
   - `input_urls`: the uploaded reference image (one URL entry; if the user uploaded several, pass only the clearest person photo);
   - `model='i2i.hd.lite'`;
   - `orientation`: match the source — `portrait` for a full-body standing shot, `square` for a square crop, `landscape` for a horizontal frame;
   - `resolution='1080P'` so face detail and hardware survive;
   - `name`: a short title such as `Leather Look - Pink Aerobics`;
   - `entity_type='characters'`.
   Never pass a physical model name; never pass width/height pixels.

6. **Verify the result against the must-keep list** (same face, same hair, same pose, same background, garment colours carried over, leather gloss visible, no extra limbs, hardware not melted, no garbled text). If it fails, run the retry ladder in `references/step-details.md` — at most two retries, then report honestly what could not be preserved.

7. **Reply briefly in the user's language.** Describe what changed (garment, gloss, hardware, colour) and what stayed the same (person, pose, scene). Do not paste long URLs or raw prompts. When the user is likely to want a next variation, call `set_guidance` in the same turn with options such as a