AI Portrait Selector & Enlarger · публичный навык Open Skill
AI Portrait Selector & Enlarger
AI Portrait Selector & Enlarger — публичный Open Skill на Digen. Откройте, чтобы увидеть, что создаёт этот процесс.
Product
1 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/rubfu8rc.jpg
# AI Portrait Selector & Enlarger
Turns a reference sheet with multiple poses (or a single uploaded portrait) into one enlarged, high-quality, photorealistic portrait of the same person. The result keeps the identity, exact pose, expression, outfit, framing and camera angle of the chosen photo, and contains only that one photo — no grid, no neighbouring thumbnails.
## When to use
- The user uploads a multi-photo collage / reference sheet and names one photo by position ("top-left", "third row second column", "the one sitting on the block") and wants it enlarged or redrawn.
- The user uploads a single portrait and asks for a high-quality, high-resolution redraw, upscale or "make it look like a real photo".
- The user asks to "pick this photo and make it bigger / cleaner / sharper" while keeping the same person and the same pose.
## Steps
1. Inspect the uploaded image(s) and the user's text. Decide whether the input is a **multi-photo sheet** (several photos of the same person in one image) or a **single photo**. If the layout is unclear, optionally call `recognize_image` to list the photos and their positions.
2. Read `references/workflow-stages.md` with `read_skill_file` and follow it for: how to identify the chosen photo, whether to cut the sheet with `split_image`, how to resolve an ambiguous or missing position, and the retry rules.
3. Obtain the selected source: the cropped tile from the sheet, or the user's single uploaded photo. Never invent a photo the user did not reference.
4. Read `references/style-and-prompt.md` with `read_skill_file` and compose the English prompt following its Style / Subject & content / Layout / Must keep / Must avoid sections and its prompt template. The prompt must describe this specific person, pose, outfit and background — do not paste a generic template.
5. Call `generate_image`:
- `prompt`: the prompt composed in step 4, in English.
- `model='i2i.hd.lite'` (image-to-image, high quality; the selected photo is the reference).
- `input_urls=[selected photo or selected tile asset]`.
- `orientation`: match the chosen photo's shape — `portrait` by default for standing/seated full-body portraits, `landscape` if the chosen photo is clearly horizontal.
- `resolution='2K'` by default for the enlargement; drop to `720P` only if the user asks for a quick draft, raise to `4K` only if the user explicitly asks for maximum size.
- `name`: a short canvas title such as `Enlarged portrait - seated on cube`.
- `entity_type='characters'`.
6. Check the result against the Must keep / Must avoid lists in `references/style-and-prompt.md` (identity, pose, outfit, framing intact; no other photos, grid lines or text in the frame). If it fails, follow the retry rules in `references/workflow-stages.md` — regenerate once with a stricter prompt, then report honestly.
7. Reply in natural language: confirm which photo was selected (by position), that identity and pose were kept, and the output size. Do not paste long URLs.
## Notes
- Only one portrait per output. If the sheet contains other people's photos, ignore them and say so.
- Every generated image is a new asset; do not claim an upscale of the original file's pixels — it is an identity-preserving redraw at high resolution.
- Content policy: refuse requests that sexualize the subject or alter a real person into a deceptive context; do not call `generate_image` for those.
- If `generate_image` returns `success=false`, tell the user the failure plainly and do not fabricate an image link.
- Keep the user's language for the conversation; the image prompt itself stays in English.