K-Pop Idol Poses from One Photo · Open Skill 公開技能

K-Pop Idol Poses from One Photo

K-Pop Idol Poses from One Photo 是 Digen 上的公開 Open Skill。開啟後即可了解此工作流能製作什麼。

@wwws

TryOn

0 0 1

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/5wp27q9f.jpg

# K-Pop Idol Poses from One Photo

Turn one uploaded photo into a set of adorable K-Pop idol pose images (9 by default) by **editing that photo**, so the person's face, hair, outfit and body proportions stay consistent across every pose. There is no structured project state: the source photo and the generated assets on the canvas are the only state.

## When to use

- The user uploads one photo and wants "K-pop idol poses", "idol photocards", "cute pose variations", "fan-edit style photos" from it.
- The user asks for a specific idol gesture from their own photo: finger heart, wink + peace sign, cheek heart, playful wave, cute salute, heart above the head, blowing a kiss, over-the-shoulder glance, hands framing the face.
- The user already got a set and wants more poses, a different expression, or one pose regenerated.
- Do **not** use it for plain text-to-image portraits with no reference photo, or for restyling a photo into a non-idol art style.

## Inputs and outputs

- **Input:** one reference photo supplied by the user (upload or an existing canvas asset). Every pose is an image-to-image edit of this photo — the photo is always passed as the reference input, never replaced by a text description.
- **Output:** images on the canvas, one per pose (9 by default), all derived from the same source photo so the person looks consistent.

## Steps

1. **Confirm the source photo.** If the user has not uploaded an image, ask for one clear photo of the person (face visible, not a collage). If several images are attached or on the canvas, ask which one is the subject, or pick the clearest single-subject photo and say which one you used. Call `recognize_image` on the chosen photo only when you need to check subject count, framing, face visibility or image quality.
2. **Track the run with `write_todos`** when generating the full set (one todo per stage: input check, pose plan, generation, quality check, delivery). Keep exactly one item `in_progress`.
3. **Read the references before writing any prompt.** Call `read_skill_file` for `references/workflow-stages.md`, then `references/style-and-prompt.md`, then `references/pose-set.md`. Use the checklist in `workflow-stages.md` to validate the photo and to run the per-image quality gate; use `pose-set.md` to choose and adapt the pose list.
4. **Plan the pose list.** Default to the 9 poses in `references/pose-set.md`. If the user asked for a specific gesture, a specific count, a different mood, or extra variations, adapt the list and state the final list in one short line before generating. If the user asks for one pose only, plan just that one.
5. **Generate one image per pose** with `generate_image`, exactly following the prompt assembly recipe in `references/style-and-prompt.md` (style block + pose block + must-keep block + must-avoid block, hard-coded in English):
   - `model='i2i.hd.lite'` — every pose is an image-to-image edit of the uploaded photo. Never pass a physical model name.
   - `input_urls=[source photo]` — the real uploaded/canvas asset of the source photo, one entry.
   - `prompt=` the assembled English prompt for that pose.
   - `name=` a short canvas title, e.g. `Idol Pose 1 - Finger Heart`.
   - `entity_type='kpop_idol_poses'` for every pose image.
   - `orientation=` match the source photo (default `portrait`; use `landscape`/`square` only if the source photo is that shape or the user asks). Do not pass `resolution` unless the user asks for higher quality (`1080P`).
6. **Quality-check each result** against the gate in `references/workflow-stages.md` (identity retained, single subject, correct hand anatomy, costume/hair unchanged, no text or watermark). Regenerate a failing pose at most twice, tightening the prompt using the fix hints for that pose; if it still fails, keep the best attempt and tell the user plainly what did not hold.
7. **Deliver.** Report in natural language which poses were produced and which ones were regenerated; do not paste long URLs. In the same tur