Beauty Filter Portrait · skill pubblica Open Skill

Beauty Filter Portrait

Beauty Filter Portrait è una Open Skill pubblica su Digen. Aprila per vedere cosa crea questo workflow.

@eggs

Other

0 0 2

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-19/np4nptsd.jpg

# Beauty Filter Portrait

Turn one uploaded photo into a beauty-filtered version of the same photo: whitened and softened skin, gently slimmed face, enlarged and brightened eyes. Output is a single retouched image that keeps the original person, pose, hair, clothing and background.

There is no structured project state. Everything you need for a shot lives in the current conversation plus the generated canvas assets.

## When to use

- The user uploads a photo and asks for a beauty / retouch / filter result on it.
- The user names one of the three effects (whitening, face slimming, eye enlargement) or all of them.
- The user wants a stronger or more natural version of a result already produced in this conversation.

Do not use this skill for: changing identity or swapping the person, changing hairstyle / outfit / background / scene, stylizing into anime or painting, video output, or adult/explicit content.

## Steps

1. **Confirm the target.** The deliverable is exactly one beauty-filtered photo of the uploaded person. If the user uploaded several photos, ask which one to retouch (or retouch the clearest front-facing one when the intent is obvious). If no photo was uploaded, ask for one — do not invent a subject.

2. **Read the source photo.** Call `recognize_image` with the uploaded image as `image_url` and a prompt asking for: face shape and feature proportions, skin tone and visible texture/blemishes, eye size and shape, eye makeup and lashes, lip shape and color, hairstyle and hair color, clothing, background, camera angle/framing, and the identity markers that must be preserved (moles, freckles placement, distinctive features, jewelry, glasses). Keep this analysis short — it feeds the prompt, it is not shown verbatim to the user.

3. **Load the visual rules.** Call `read_skill_file` for `references/style-and-prompt.md` and `references/step-details.md`. Follow the prompt template, the keep/avoid lists and the intensity ladder there. Do not write the prompt from memory.

4. **Write the prompt in English.** Fill the template from `references/style-and-prompt.md` with the details from step 2:
   - subject block = the actual person described in the uploaded photo (face shape, hair, clothing, background, lighting),
   - retouch block = the requested effects at the intensity level chosen in `references/step-details.md`,
   - keep block = the identity markers captured in step 2,
   - avoid block = the must-avoid list from the reference file.
   The prompt must be specific to this photo; a generic "beautiful woman, retouched" prompt is not acceptable.

5. **Generate the image.** Call `generate_image` with:
   - `prompt` = the prompt from step 4,
   - `input_urls` = [the uploaded photo],
   - `model='i2i.hd.lite'`,
   - `orientation` matched to the input photo (portrait for vertical selfies, landscape for horizontal, square for square) — see `references/step-details.md`,
   - `resolution='1080P'` when the user wants a sharp close-up result; otherwise omit it and use the default,
   - `name` = a short label such as "Beauty filter result",
   - `entity_type='beauty_portraits'`.

6. **Branch only if needed.** If the user explicitly asks for a from-scratch beauty portrait with no photo (or refuses to upload one), use the text-to-image branch in `references/step-details.md` with `model='t2i'`. If the user asks for a before/after comparison card, follow the card recipe in the same file.

7. **Reply.** Describe in one or two sentences what was retouched (skin brightness, jawline, eyes) and offer to go stronger or more natural. Do not paste long URLs.

## Notes

- Identity is non-negotiable: same person, same facial structure, same pose, same crop, same hair, same clothing, same background. Beauty filtering is a retouch, not a recast.
- Never over-retouch: no plastic/doll skin, no washed-out white, no obviously enlarged eyes or cartoon proportions. Stay inside the limits in `references/style-and-prompt.md`.
- The output is a clean photo. D