Couple Photo Generator · Open Skill 公開技能
Couple Photo Generator
Couple Photo Generator 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。
Custom
3 0 3
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284862/2026-10-16/afr7ntmp.jpg
# Couple Photo Generator Turn the user's uploaded reference photos into one joint couple photo: two people, one frame, both faces staying themselves, in the single unified "5 couple photo" editorial look. There is no structured project state — everything lives in this conversation, the generated canvas assets, and todos. ## When to use - The user uploads (or points at) photo(s) of two people and asks for a couple photo / their picture together / a couple-photo-style joint portrait. - The user wants the same two people re-shot with a different pose, background, light or crop, while keeping their identities. - Not for: group photos of three or more people, single-person headshots, turning a person into someone else, or non-photo art styles. ## Inputs and outputs - Input: **reference image(s)** — usually one or two photos of real people; each person must be present in at least one reference. - Output: **one image**, added to the canvas as `entity_type='couple_photos'`. ## Steps 1. Before composing anything, read the visual conventions: `read_skill_file(name="couple-photo-generator", path="references/style-and-prompt.md")`. The Style / Subject / Layout / Must keep / Must avoid rules and the English prompt template all live there. Do not write a prompt before reading it. 2. Establish the two people and their reference URLs: - one reference containing both people → use that single image; - two separate portraits → map them to Person A and Person B explicitly, and use one URL per person in `input_urls` (order: Person A first, Person B second), naming each in the prompt; - only one of the two people is available → ask the user for the second person's photo. Never invent, borrow or stylise a missing face, and never substitute a public figure. 3. Write the image prompt in **English**, following the template in `references/style-and-prompt.md`: list the observable identity attributes to keep for each person (face shape, hairstyle, hair colour, glasses, facial hair, skin tone, clothing colour and material), then the unified style block, then the requested framing/background/pose. If the user asked for changes, phrase them as variations of the same style block — never as a different look. 4. Call `generate_image` with: - `prompt`: the English prompt from step 3; - `input_urls`: the reference URL(s) from step 2 (image-to-image, not text-to-image); - `model='i2i.hd.lite'` (image-to-image channel; reference photos are always supplied); - `orientation='portrait'` by default — use `'landscape'` only when the user asks for a horizontal shot, `'square'` when they want a social-profile crop; - `resolution='2K'` for the magazine-grade look; the platform may auto-downgrade to the channel's ceiling and report it in the tool result; - a short `name` (e.g. "Couple portrait — park afternoon") and `entity_type='couple_photos'`. 5. Reply in the user's language with one or two sentences on what was produced. Do not paste long URLs. If the user wants more than one frame, repeat step 4 with a small variation in pose/background/lighting while keeping the identity block identical — the platform generates one image per call, so never ask for a batch. ## Notes - Identity is the top priority: both faces must stay the people in the references. If the user says a face drifted, re-read `references/style-and-prompt.md`, strengthen the "Must keep" wording, and regenerate with the clearest available reference per person. - Keep the style unified: palette, materials and light direction are the same for both people and across the whole frame. Never mix looks inside one image. - `success=false` → report the failure honestly; never claim a result exists or make up a URL. - Do not add text, watermarks, logos, captions, extra people or collage/grid effects that the user did not ask for. - Refuse sexual, nude or otherwise adult content involving real people, and refuse requests to pair someone with a non-consenting or fabricated ident