Crosswalk Motion Portrait · compétence Open Skill
Crosswalk Motion Portrait
Crosswalk Motion Portrait est une Open Skill publique sur Digen. Ouvrez-la pour voir ce que ce workflow crée.
Sports
0 0 0
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4150474/2026-10-15/hcr5ckaq.jpg
# Crosswalk Motion Portrait
Generate a **vertical 3:4 cinematic street portrait** from the user's text (optionally with 1–2 reference images): the subject stands perfectly still and keeps the same face and identity as the uploaded person, while everyone around them is smeared into motion trails.
The whole visual formula is one sentence: **the subject's time runs at normal speed; everyone else's time is stretched out.**
## When to use
- The user wants to generate or draw this kind of "street corner / crosswalk + crowd trails" portrait.
- The user provides a photo of a person and wants it restyled into this film-still look (identity must be preserved).
- The user provides a street / crosswalk photo to use as the composition base.
- The user wants to change a single thing (outfit, flower, city, pose) and regenerate the same look.
## Inputs and how the two reference images are used
| Reference image | Purpose | Where it goes |
|---|---|---|
| Person reference (1st image) | Locks identity: face shape, facial features, hairstyle and hair colour, age, body type | `input_urls[0]` |
| Crosswalk scene reference (2nd image) | Locks the composition base: street, storefronts, signage, perspective, road surface | `input_urls[1]` |
- Both provided: `input_urls=[person image, scene image]` (order fixed, person first).
- Only the person image: generate the scene from the user's text theme.
- Only the scene image: generate the person from the user's text description, with no identity lock.
- Neither: pure text-to-image (`model='t2i'`).
- Images uploaded by the user can be passed through directly using their asset reference — **never** invent URLs or image links.
## Steps
1. **Decide the mode and theme slots.** Pin down the requested outfit `{outfit}`, prop, location `{location}` and city elements. If the user says nothing, use the default theme: brown oversized wool suit + pale pink tulip + London-style street corner (Monmouth-style coffee shop signage, red double-decker bus, yellow traffic lights, wet crosswalk).
2. **Read the prompt recipe**: call `read_skill_file(name='crosswalk-motion-portrait', path='references/style-prompt.md')` and assemble the English prompt using the layered skeleton in it. **Do not improvise from memory** — the way the trails are written, the lighting, the grade and the lens parameters all live in that file.
3. **Call the generation tool** with fixed parameters:
- With reference images: `generate_image(prompt=<assembled English prompt>, input_urls=[person image, scene image], model='i2i', orientation='portrait', aspect_ratio='3:4', name=<short title, e.g. "Stillness in the Crosswalk">, entity_type='portraits')`
- Without reference images: same as above, but drop `input_urls` and use `model='t2i'`
- `aspect_ratio='3:4'` is the hard convention for this layout (vertical 3:4); do not change it to 16:9 or 9:16; pass `orientation='portrait'` together with it.
- `name` should be a short title for the subject in the image (Chinese or English both fine) so the canvas can display it.
- Do not pass `group_id` / `index` / pixel-size parameters; if no specific resolution is requested, do not pass `resolution` (defaults to 720P).
4. **Check the result.** Verify three things: the subject is crisp and sharp, the surrounding people have clear motion trails, and the grade is unified and desaturated. If not satisfied, retry once with concrete feedback (e.g. "the subject is blurred too", "the trails are not long enough"); do not blindly retry more than 2 times.
5. **Reply.** Summarize in one or two natural-language sentences what was made (theme, whether identity was preserved, the trail effect); **do not paste long URLs** — the asset goes to the canvas automatically. You may also ask whether they want a different theme for another image.
## Notes
- **Identity first**: with a reference image, the face and overall identity must remain the same person; only change outfit, pose and scene — do not "beautify" or chan