Reference Photo Studio r1–r5 · Open Skill 공개 스킬
Reference Photo Studio r1–r5
Reference Photo Studio r1–r5은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.
Sports
5 0 5
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-19/zrd6mwkp.jpg
--- allowed-tools: - generate_image - recognize_image --- # Reference Photo Studio r1–r5 Turns one to five uploaded reference photos into a single realistic, highly detailed studio photograph of those same people, with the interaction the user describes. The uploaded images are addressed as `r1`–`r5` inside the generation prompt. Output is one image — no video, no structured project state. ## When to use - The user uploads 1–5 photos of people and asks for a new photo of them together (couple, friends, family, cast). - The user asks to place the uploaded people in a specific interaction, pose, or studio setting. - The user wants an existing photo set re-composed or restyled into one consistent studio frame. - The user only has text and no reference image: ask for the photos first — this skill is reference-driven, not text-only. ## Inputs and outputs - **Input — reference images (required):** assign labels `r1`…`r5` in upload order. Each label must be used verbatim in the prompt and passed to the tool in the same order. - **Input — text (optional but typical):** who appears, the interaction (gaze, touch, distance), setting, wardrobe, crop, mood. Everything not stated is inferred from the references, not invented. - **Output:** exactly one image via `generate_image` with `model='i2i.hd.lite'`. ## Steps 1. **Inventory the inputs.** Label the uploaded images `r1`, `r2`, … in the order they were given (max 5; if more, keep the first five and tell the user which ones you used). Then read `references/reference-mapping.md` with `read_skill_file` and follow its labelling, inspection, and layout-branch rules. 2. **Parse the request.** Extract: which labels participate, the interaction, the setting, wardrobe, crop, and mood. If the interaction is ambiguous, choose the most conservative plausible reading (e.g. standing side by side) and state that choice — do not stall for confirmation. 3. **Inspect each reference before writing the prompt.** Call `recognize_image` per reference only when you cannot confidently describe that person (identity cues: face shape, hair colour and texture, eye colour, skin tone, build, wardrobe, accessories). Record one short line per label for your own prompt writing; never invent features that contradict the photo. 4. **Read the style contract.** Call `read_skill_file` for `references/style-and-prompt.md`. It defines Style, Subject/content, Layout, Must keep, Must avoid, the prompt template, and copy-ready style fragments. 5. **Read the prompt recipes.** Call `read_skill_file` for `references/prompt-recipes.md` and pick the recipe that matches the layout (solo / pair / group). Fill its slots with the actual people from step 3 and the interaction from step 2. The result must be a specific English prompt naming `r1`…`r5` explicitly — never a generic "beautiful studio portrait" sentence. 6. **Generate.** Call `generate_image` with: - `prompt`: the finished English prompt from step 5, - `input_urls`: the reference images in label order (`r1` first), - `model='i2i.hd.lite'`, - `orientation`: `'portrait'` by default; `'landscape'` or `'square'` only when the user asks or the described composition requires it, - `resolution`: `'1080P'` (raise to `'2K'` only if the user asks for print/high detail), - `name`: a short human title, e.g. `Studio Portrait r1-r2`, - `entity_type='characters'`. 7. **Reply and offer the next move.** Summarise in one or two sentences which references were used and what was composed; do not paste URLs. In the same turn call `set_guidance` with stage-relevant options (adjust interaction, change backdrop, add another reference, switch crop). ## Notes - Identity fidelity comes first: keep each person's face, hair, skin tone, build, and age exactly as in their reference; the prompt may describe wardrobe and pose changes only when the user asks for them. - Only include people who exist in the references. Never add extra characters, text, logos, labels, or the ref