Lingerie Outfit Filter #17 · Open Skill 公开技能
Lingerie Outfit Filter #17
Lingerie Outfit Filter #17 是 Digen 上的公开 Open Skill。打开后即可了解该工作流能制作什么。
Doc
10 6 11
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/wgf78m2m.jpg
# Lingerie Outfit Filter #17 — Photo to Photo Outfit Swap
This skill reproduces the single-purpose app "Lingerie Filter #17": it takes **one photo of a person** and re-dresses that same person in the **red lace babydoll / harness lingerie set** from the app cover. It is a one-shot image-to-image filter — there is no project state and no multi-stage pipeline. Read the reference files, build one prompt, call `generate_image` once, report back in plain language.
## When to use
- The user uploads a photo of a woman and asks for the lingerie filter / for her clothes to be replaced with the outfit on the cover.
- Works for real people and for animated, 3D or illustrated characters (the source art style is preserved).
- The user asks for the same outfit in a different colour ("color selector": black, pink, purple, white, etc.).
- The user wants a "lingerie photoshoot" version of an existing photo.
Do not use this skill when there is no subject photo: ask the user to upload one instead of inventing a person.
## Inputs and output
| | What | How it is used |
|---|---|---|
| Input (required) | Subject photo — the person to re-dress | `input_urls[0]`: asset_id / `s3://` uri / https URL |
| Input (optional) | Garment reference — the cover image of the outfit | `input_urls[1]` when available; otherwise describe the outfit from `references/garment-spec.md` |
| Output | One still image, same framing as the source | generated asset on the canvas |
## Steps
1. Identify the subject photo and confirm what must stay untouched (face, pose, background). If several photos are attached, ask which person to dress; if none, ask for one.
2. Run the pre-flight checklist in `references/step-details.md` (adult subject, source quality, colour request, keep-list). Stop and refuse if it fails.
3. Before generating, call `read_skill_file('lingerie-outfit-filter', 'references/style-and-prompt.md')` and `read_skill_file('lingerie-outfit-filter', 'references/garment-spec.md')`. Read `references/step-details.md` too when you need the branch rules, the colour-variant handling or the retry table. Do not generate before reading them.
4. Build one English prompt from the template and rules in `references/style-and-prompt.md`, inserting the garment description from `references/garment-spec.md` (or relying on the cover image at `input_urls[1]`). State explicitly that image 1 supplies identity and image 2 supplies the outfit.
5. Call `generate_image` with the agreed parameters:
- `model='i2i'` — the bare image-to-image channel. Never a `.hd` / `.hd.lite` variant and never a physical model name; the higher-quality variants reject this kind of suggestive content.
- `input_urls=[subject_photo]` or `input_urls=[subject_photo, garment_reference]`.
- `prompt` — follow `references/style-and-prompt.md`.
- `orientation` — match the source photo; default `portrait` when unsure. `resolution` — 720P by default, 1080P only if the user asks for extra detail.
- `name` — a short title such as `Lingerie filter portrait`.
- `entity_type='characters'` — the output stays a portrait of the same person.
6. Check the result against the post-generation checklist in `references/step-details.md` (identity kept, only clothing changed, all garment elements present, no nudity, no added text or props). If one item fails, retry once with the targeted fix from the retry table; at most two attempts total.
7. Reply in one or two natural sentences describing what changed. Do not paste long asset URLs. If it is useful, call `set_guidance` in the same turn with two or three concrete next steps (another colour, a close-up, a different photo).
## Parameter conventions
- Channel is fixed: `model='i2i'` for every generation in this skill.
- The number of `input_urls` decides the branch: 2 = photo + cover image (preferred, most faithful); 1 = photo only, outfit rebuilt from the written spec.
- Never pass `group_id` / `index`; never pass `width` / `height` — only `orientation` and `reso