ACO Video Director Pro 2.8 · Open Skill 공개 스킬
ACO Video Director Pro 2.8
ACO Video Director Pro 2.8은 Digen의 공개 Open Skill입니다. 열어 보면 이 워크플로가 무엇을 만드는지 알 수 있습니다.
Film
655 37 292
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/dev/admin/100/2026-08-19/9eagj9ab.jpg
--- allowed-tools: - generate_image - generate_video - generate_tts - mix_clips - transcribe_audio - recognize_image - recognize_video - create_document - update_document - get_library_item - list_models - memory_whoami - memory_query - memory_store --- # ACO Video Director Pro 2.8 ## When to use - The user wants a music video built around an uploaded song — audio-driven, tight lip-sync, entire runtime of the song. - The user wants a cinematic video (movie, short film, series episode, feature) focused on story, with no music background. - The user wants any-length video by generating 10–15s clips and stitching them together into videos of unlimited length. - The user wants characters kept consistent using uploaded reference images, with a character sheet per character. - The user wants storyboards rendered and approved before any video is generated. - The user wants to advance or revise only one stage of such a production. - The user wants to run the stringent QC/likeness protocol against an existing video, keyframe, or clip. ## Core rules (non-negotiable) 1. **Ask first: movie or music video?** Determine the mode before anything else. 2. **Ask for a visual style, listing every available style**, and let the user pick one (default **ultra realistic** unless they choose another). When presenting the list, also **recognize and note which styles pair well** with each other (see "Visual styles & pairing" below) so the user can combine compatible looks. 3. **Present HDR as an option at the start.** When the skill is chosen, ask the user whether to enable HDR (default: **on** when available). Present it as a clear choice alongside the visual-style selection, before any generation begins. If the user opts out, use the best non-HDR tier. 4. **Ask for the aspect ratio.** Present the choice up front: **portrait 9:16** (vertical, default for short-form / social), **landscape 16:9** (widescreen), **square 1:1**, or a **custom ratio** (e.g. 3:4, 4:3, 2:1). Apply the chosen ratio consistently to every keyframe, clip, and the final cut (pass `orientation` and/or `aspect_ratio` accordingly). Default to **portrait 9:16** unless the user chooses otherwise. 5. **No AI song generation.** In music video mode the soundtrack is always the uploaded song. Never call `generate_music`. 6. **Music video mode uses the entire runtime of the song.** The song plays in full, never shortened, never looped; the finished video must be at least as long as the song. 7. **Music video is audio-driven with tight lip-sync.** Transcribe the uploaded song with `transcribe_audio`; the words and their timing drive the lip-sync, and the visuals tightly follow the story in the song. 8. **Every shot maps the song.** Detect the hook and verses and plan shots so each maps to a specific song section, in order, covering the whole song. 9. **Cinematic video has no music background.** It focuses on story; dialogue and sound design carry it. 10. **Generate a character sheet for each uploaded character** (appearance, wardrobe, personality, role, reference image) before any visuals. 11. **Create a script or work from an uploaded script/plot.** Approve the script or story before storyboards. 12. **Create keyframes** for every shot (start/end frames) so seams are exact. 13. **Each shot is 10–15 seconds.** 14. **Video stitching creates videos of unlimited length.** 15. **ALWAYS use the character reference image(s) in every generation** — every storyboard frame and every clip must be generated with `input_urls` containing the relevant character reference(s). Never generate a character from text alone when a reference exists. This is non-negotiable. 16. **Blend multiple references when a character asset exists.** When a character has more than one reference image (e.g. a full-body shot plus face shots), pass ALL of them together in `input_urls` so the model blends them into one consistent likeness. Do not rely on a single cropped face or a single full body