Fashion Runway Walk AI · Open Skill 公开技能
Fashion Runway Walk AI
Fashion Runway Walk AI 是 Digen 上的公开 Open Skill。打开后即可了解该工作流能制作什么。
TryOn
1 0 3
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-21/gnbuwc9h.jpg
# Fashion Runway Walk AI Turns one uploaded portrait into a short vertical catwalk clip: the same person (same face, same identity) walking a glossy runway in a themed outfit, followed from a low camera angle, audience blurred on both sides. There is no structured project state — the uploaded portrait plus the assets already on the canvas are the state. ## When to use - The user uploads a photo of a person and asks for a "runway walk", "catwalk", "fashion show", "model walk" video. - The user wants that same clip restyled: different outfit / theme / season / color story, or a different camera framing. - The user wants only the styled still (a runway look image) — then do steps 1–5 and stop. ## Inputs and outputs | | | |---|---| | Input | One reference portrait (required). A second reference image (outfit, runway, or mood reference) is optional. | | Output | A short vertical (9:16) runway-walk video, plus one styled look still used as the visual anchor. | How the input is used in prompts: - **The user's portrait carries identity only.** In every prompt, describe it as "the same person as the reference: identical facial features, face shape, skin tone, hair color and hairline". Never re-describe the face from scratch, never invent a different person. - **The optional second reference carries wardrobe/scene only.** Take its garment silhouette, fabric and color, and its runway setting; never take its face. - **Everything else (pose, garment, scene, lighting, camera) comes from this Skill's reference files**, not from guesswork. ## Steps 1. Look at the uploaded portrait with `recognize_image` (`prompt` = ask for age range, gender presentation, hair color/length, skin tone, face shape, current clothing, framing, sharpness). This description is what you will reuse as the identity anchor. 2. Read `references/step-details.md` and run its intake checklist: single person, face clearly visible, no heavy blur, no occlusion. If it fails, ask the user for a clearer portrait and stop. 3. Confirm the look in one short message: theme (default: clean white mini-dress runway look), and whether they want the clip or only the still. Do not ask more than one round of questions unless the request is genuinely ambiguous. 4. Read `references/style-and-prompt.md`. Following its Style / Subject / Layout / Must-keep / Must-avoid rules, write the **look-still prompt** for this specific user (their theme, garment, color story) — do not paste the template verbatim. 5. Call `generate_image` with `model='i2i.hd.lite'`, `input_urls=[the user's portrait]`, `prompt=<the look-still prompt>`, `orientation='portrait'`, `resolution='1080P'`, `name='<theme> runway look'`, `entity_type='characters'`. If the user only wanted a still image, reply and stop here. 6. Read `references/video-script.md` and write a **full, detailed English script for this exact request** — logline, duration, beats with time codes, shot size, camera movement, subject action, garment/hair behaviour, lighting and atmosphere, transition, must-keep, must-avoid. Fill it from the user's own words plus the identity anchor from step 1. A generic template is not an acceptable script. 7. Read `references/video-content-constraints.md` and check your script against its shot/motion/prohibition rules; fix the script before going on. 8. Compress the finished script into the video prompt (identity + garment + action + camera + environment + lighting + prohibitions) and call `generate_video` with `model='ref2v'`, `input_urls=[the user's portrait, the look still from step 5]`, `prompt=<compressed script>`, `orientation='portrait'`, `resolution='1080P'`, `duration` as decided in the script, `name='<theme> runway walk'`, `entity_type='final_videos'`. 9. Optionally verify the result with `recognize_video` (`prompt` = "is this the same person as the reference portrait, and is the garment/runway consistent?"). If identity drifted or the garment flickered, regenerate once with the drift named explicitly in th