Couple Portrait Studio · Open Skill 公开技能
Couple Portrait Studio
Couple Portrait Studio 是 Digen 上的公开 Open Skill。打开后即可了解该工作流能制作什么。
Auto
1 0 2
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-19/kdwqazqy.jpg
# Couple Portrait Studio
Turns one (or two) uploaded reference photos into photorealistic romantic couple portraits: the person from the upload stays exactly who they are, and a believable partner is placed in the same frame inside a warm, dreamy, golden-hour scene. There is no structured project state — the reference images, the identity notes you write down, and the generated assets on the canvas are the whole state.
## When to use
- The user uploads a photo and asks for a couple photo, "us together", a boyfriend/girlfriend portrait, a matching couple picture, or a romantic photo with a partner.
- The user wants that same couple photo repeated in several scenes ("give me a set", "14 photos", "different backgrounds").
- The user wants an existing couple portrait re-shot with a different pose, scene, wardrobe, or mood.
- Do not use for solo portraits with no partner, product shots, or poster/UI design work.
## Inputs and outputs
- Input: one reference photo (the user's own subject) — required for the main path. Optionally a second photo of the partner; if given, both identities are preserved and blended into one frame.
- Output: images only (vertical portrait photos). Assets go to the canvas with `entity_type='couple_portraits'`.
## Steps
1. **Read the request.** Identify: who is in the uploaded photo, whether a second person's photo was supplied, which scene/mood the user asked for (or "let the skill choose"), and how many images they want. If no reference photo is attached, ask for one — a couple portrait must be built on the real subject. Only if the user explicitly refuses to upload anything, fall back to text-only generation with `model='t2i'` and describe the subject from their words.
2. **Read `references/step-details.md`** with `read_skill_file` and follow its intake checklist, identity-extraction rules, partner-design rules, set-size rules, and QC/retry rules.
3. **Inspect the reference image(s).** Call `recognize_image` on each uploaded photo with a prompt asking for an exact physical description: face shape, eyes, brows, nose, mouth, jawline, skin tone, hairstyle and hairline, facial hair, glasses, moles/marks, apparent age, build, and the clothing visible in the shot. Write this down as the identity lock; it is what keeps the person recognizable.
4. **Read the visual rules.** Call `read_skill_file` for `references/style-and-prompt.md` (style, subject handling, layout, must-keep, must-avoid, prompt template) and, when the user wants a scene or a set, `references/scene-variants.md` (the fourteen scene recipes with pose, wardrobe, light and location ready to paste).
5. **Write the English prompt** for the current image using the template from `references/style-and-prompt.md`: fill the identity lock from step 3, the scene/pose/wardrobe/light slots from the user's request or from `references/scene-variants.md`, and keep the must-keep and must-avoid clauses verbatim. Never ship a generic template sentence as the final prompt — every slot must be specific to this user and this scene.
6. **Generate with the image-to-image channel.** Call `generate_image` with `model='i2i.hd.lite'`, `input_urls=[reference_url]` (both photos when a partner photo was supplied, first URL = the user's subject), `orientation='portrait'`, `resolution='2K'`, a short descriptive `name` (e.g. `Couple Portrait - Golden Hour Park`), and `entity_type='couple_portraits'`. Add `aspect_ratio='3:4'` only if the user asks for a classic photo-print ratio.
7. **Build a set one image at a time.** The tool returns one image per call; for a multi-photo set, repeat steps 5–6 once per scene with the same identity lock, changing only scene/pose/wardrobe/light, and vary the camera distance between shots (close two-shot, medium, wider environmental). Keep the default to 4 images unless the user asks for the full fourteen-scene set, and never exceed fourteen.
8. **Check each result** against the QC list in `references/step-details.md` (identity match, hands, ey