IdolDance AI · Open Skill 公开技能
IdolDance AI
IdolDance 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
0 1 3
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4284865/2026-10-20/jzf492ej.jpg
# IdolDance AI Turn one uploaded portrait into a short vertical dance video: the same person, restyled as an idol performing on a glossy stage. There is no structured project state — the reference image plus the canvas assets are the only state, and progress is tracked in conversation and `write_todos`. ## When to use - The user uploads a portrait and says "make this dance", "turn this into a dance video", "idol dance", "K-pop dance challenge", or similar. - The user points at an image already on the canvas (previous generation or upload) and wants a dance clip from it. - The user wants an existing dance clip redone or adjusted: different outfit, different stage, calmer or wilder choreography, closer or wider framing. Do not use this skill for stills only (just describe or hand off to an image skill), for videos of a person that is *not* taken from a user image, or for multi-shot edited movies (that is an editing task, not a single generated shot). ## Input and output - Input: exactly one reference image — a portrait photo of the person who should dance. The person's face, facial features, skin tone, hairline, and body proportions come from this image and must be preserved. - Output: one short vertical dance video produced by a single `generate_video` call with `model='ref2v'`. - The reference image is passed as `input_urls=[<image asset id or s3:// uri or https url>]`. Do not paste image URLs into the reply text; refer to the reference as "your photo". ## Steps 1. **Confirm the reference image.** Read `references/step-details.md` first and follow its input checklist and branch rules (no image / group photo / not a person / too blurry). If the user gave no usable image, ask for one clear portrait photo and stop — never invent a person or reuse an unrelated asset. 2. **Analyze the reference image** with `recognize_image` (short prompt asking for: apparent gender presentation, age range, hair, face-shape and distinctive features, current clothing, current framing and background, and whether it is a photo or an illustration). Keep the result as the identity description used later. 3. **Plan the work.** For anything beyond a single one-off clip call `write_todos` with the stages: analyze photo → write script → build prompt → generate video → deliver. 4. **Read the visual rules**: `read_skill_file(name="idol-dance-ai", path="references/style-and-prompt.md")`. Use its Style / Content / Layout / Must keep / Must avoid sections and its prompt template to define the look for this request. 5. **Write this request's script.** `read_skill_file(name="idol-dance-ai", path="references/video-script.md")`, then write a full English script for *this* user request following that file's structure (logline, runtime and ratio, identity lock, beat sheet with shot size / camera / body action / lighting / transition, continuity, must-keep, must-avoid). It must follow the user's stated theme, mood, outfit idea, and energy — a generic template is not acceptable as the finished script. If the user only wants a still image, skip this step and the video step. 6. **Read the shot limits**: `read_skill_file(name="idol-dance-ai", path="references/video-content-constraints.md")`. Apply its single-shot, camera-motion, choreography-readability, and forbidden-content rules while compressing the script. 7. **Compose the video prompt** from the script plus the style rules: identity lock sentence first, then wardrobe and stage theme, then the beat-by-beat action, then camera behavior, then lighting and grade, then the negative list. Keep it a dense paragraph (roughly 120–200 words), not a bullet dump. 8. **Generate the video.** Call: `generate_video(prompt=<composed prompt>, input_urls=[<reference image>], model='ref2v', orientation='portrait', duration=5, name="Idol Dance – <short theme tag>", entity_type='final_videos')` Conventions: `model='ref2v'` (reference-based video, the channel for image-driven dance clips — never a physical model name, never