Bathroom Scene Transform · skill pública Open Skill

Bathroom Scene Transform

Bathroom Scene Transform é uma Open Skill pública no Digen. Abra para ver o que este fluxo cria.

@wwws

Product

2 0 0

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/md336mgg.jpg

# Bathroom Scene Transform

Turn one uploaded photo of a person into a warm, tasteful bathtub portrait: the person's identity stays exactly the same, while wardrobe, pose and setting are rebuilt as a bathroom bath scene. Single-image, image-to-image task; no structured state is kept between turns — later revisions simply reuse the photo already in this conversation or the asset just generated.

## When to use

- The user uploads (or already uploaded) a photo of a person and asks for a bathroom / bathtub / bath-time version of it, a bathroom background replacement, or a "bath scene photo".
- The user asks for a variant of a bath portrait produced earlier in this conversation (different light, different pose, towel instead of robe, another angle).
- Do **not** use when there is no person photo to edit (text-only idea for a bath scene): ask for the photo once, then wait. Do not use for non-bathroom restyling requests.

## Inputs and outputs

| Input | How it is used |
|---|---|
| Reference photo (required) | The person to keep. Passed to `generate_image` as `input_urls=[<photo>]`; its identity cues are restated in the prompt's identity-lock clause. |
| Text (optional) | Style modifiers only: light (morning / evening lamp), color mood, framing, wardrobe, pose, prop details, or revision notes. Folded into the prompt slots; never used to change the person's identity or to break content policy. |
| Output | One bathroom-scene portrait image, added to the canvas as `entity_type='characters'`. |

## Steps

1. **Check the photo.** Use the photo the user uploaded in this conversation (agent can reference it by its asset / `s3://` / https value). If there is no photo yet, ask for exactly one clear photo of one person in a single short message and stop. If there are several photos, pick the clearest single-person, well-lit one and say which one you used.
2. **Read `references/style-and-prompt.md`** with `read_skill_file` before writing any prompt. It defines the fixed style (color, materials, light), the subject/content rules, the layout, the must-keep and must-avoid lists, and the copyable English prompt template plus fragment library.
3. **Optionally lock identity.** When the photo is small, angled, partly covered, or the face is unclear, call `recognize_image(image_url=<photo>, prompt="Describe only identity-critical visual details: apparent age range, gender presentation, face shape, eye shape, nose and mouth shape, hairstyle, hair color and length, skin tone, build, and current clothing.")`. Use the answer only to strengthen the identity clause — never to change who the person is, and never to guess sensitive attributes.
4. **Read `references/step-details.md`** when you need the prompt assembly order, the parameter-choice rules, group-photo handling, conflict handling for user text, or the QC / retry checklist.
5. **Write the English prompt** by filling the template from `references/style-and-prompt.md`: identity-lock clause first, then wardrobe, pose, setting, warm bathroom light, camera/lens, mood, and the trailing avoid clause. Keep it roughly 60–130 words. Fold the user's text into those slots only; silently drop anything that would alter the identity or the fixed theme, and refuse anything that breaks content policy.
6. **Generate** with the image-to-image channel:
   `generate_image(prompt=<composed prompt>, input_urls=[<uploaded photo>], model='i2i.hd.lite', orientation=<match the uploaded photo — portrait / landscape / square; default 'portrait'>, resolution='1080P', name='Bathroom Portrait', entity_type='characters')`.
   Never pass a physical model name; always `model='i2i.hd.lite'` for this photo-to-photo task. Do not pass `aspect_ratio` unless the user explicitly names a non-standard ratio.
7. **Quality check, then at most one retry.** Run the checklist in `references/step-details.md`: same person as the upload, bathroom scene actually readable, one subject, tasteful full coverage, no text or watermark, believable hands an