Keep Second Image Pose · Open Skill 공개 스킬

Keep Second Image Pose

Keep Second Image Pose은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.

@wwws

Other

2 0 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284863/2026-10-20/muypc59y.jpg

---

allowed-tools:
  - recognize_image
  - generate_image
---

# Keep Second Image Pose

Turn two uploaded reference photos into one new picture: the person (or subject) from the first image, placed into the body position of the second image, rendered as a single coherent photograph. There is no structured project state — everything lives in the conversation and the canvas assets.

## When to use

- The user uploads an image (or two) and asks to "keep the pose of the second picture", "same posture as this reference", "put this pose on my character".
- The user wants a specific body position copied onto a subject: seated, legs raised, leaning against a wall, arms crossed at the waist, and so on.
- The user asks to redo or repair a previous pose-transfer result (limbs wrong, pose drifted, identity changed).
- Do not use this skill for pure text-to-image requests with no reference photo, and not for editing the scene/environment while keeping the pose unrelated to the request.

## Inputs and outputs

- Input: reference images. Image 1 = **subject source** (identity, face, hair, clothing, body); Image 2 = **pose source** (body position, support surfaces, camera angle). If the user names them differently, follow the user.
- Output: one generated image on the canvas. Channel is always `i2i.hd.lite` (image-to-image with references), never a text-to-image call.

## Steps

1. Read the request and identify the images. If the user attached two images, assign image 1 = subject, image 2 = pose. If only one image is attached and the pose must come from a second picture, ask for it briefly; if the user insists on a written pose instead, keep the single image as subject and describe the pose in words.
2. Call `read_skill_file` for `references/style-and-prompt.md` and follow its Style / Subject / Layout / Must-keep / Must-avoid rules. All prompt wording for this skill comes from that file.
3. Call `read_skill_file` for `references/step-details.md` and run its per-step checks: image-role triage, pose lock extraction, prompt assembly order, generation parameters, and the post-generation QA loop.
4. Analyze both images with `recognize_image` (one call per image) using the extraction prompts in `step-details.md`, so the pose (angles, contact points, chair/wall geometry) and the subject (face, hair, garment list and colors/prints) are described precisely instead of guessed.
5. Assemble one English `prompt` from the template in `references/style-and-prompt.md`: subject paragraph (from image 1), pose paragraph (from image 2), set/lighting paragraph, then the unified style and framing sentences. Keep one single visual theme for the whole frame.
6. Call `generate_image`:
   - `prompt` = the assembled prompt,
   - `input_urls` = `[subject_image, pose_image]` in that exact order (omit the second entry only in the single-image fallback),
   - `model='i2i.hd.lite'`,
   - `orientation` from the pose image aspect (`portrait` for vertical standing/raised-leg poses like the reference, `landscape` if the pose reference is horizontal, default `portrait` when unsure),
   - `resolution='2K'` when the user asks for high clarity or a printable frame, otherwise leave it unset,
   - `name` = short title such as "Pose transfer – seated legs up",
   - `entity_type='posed_images'`.
7. Run the QA comparison described in `step-details.md` (`recognize_image` on the result, checklist of identity / pose / contacts / limbs / framing). Retry at most once by naming the failed item in the prompt; after that, report honestly what still differs.
8. Reply in plain language with what was produced and which details were preserved. Do not paste long URLs. Call `set_guidance` in the same turn with 2–4 relevant next actions (e.g. strengthen the pose, swap the outfit, vertical crop).

## Notes

- Never change the subject's identity, hair, garment set, prints, or the pose's contact points unless the user explicitly asks.
- Do not invent app features, controls, or presets beyond pose trans