Cinematic 3D Video Recreation · Open Skill 公開スキル
Cinematic 3D Video Recreation
Cinematic 3D Video Recreation は Digen 上の公開 Open Skill です。開くと、このワークフローが何を作るか分かります。
Anime
9 2 5
https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4163347/2026-09-24/eq65yzp2.png
# Cinematic 3D Video Recreation Recreate a supplied reference video as an exceptionally polished, cinematic 3D animation — same story, same visual progression, same emotional tone — with characters that stay identical in every shot. ## When to use - The user provides a reference video (uploaded asset or a social link) and wants it recreated as a high-end 3D animated video. - The user wants to refine a single shot, restyle a character, or re-mix the final cut of such a recreation. ## Reference files - `references/style-guide.md`: the default premium 3D family-animation style, character-design defaults, environment, lighting, quality/format specs and prohibitions. Read it before generating anything, and apply it unless the reference video clearly dictates otherwise. ## Steps 1. **Get the reference video.** If the user uploaded it, use its asset URL directly. If it's a YouTube/TikTok/Instagram link, call `download_video(url=...)` and use the returned `downloadUrl` (pass it verbatim, even if it looks like `shorturl://…`). For Facebook/X/LinkedIn links, `download_video` is unsupported — call `fetch_video_info(url=...)` for metadata only and tell the user we can't fetch the file. 2. **Analyze the reference.** Call `recognize_video(video_url=..., prompt=...)` to extract: shot-by-shot breakdown, composition and framing, camera language, character behavior and motion, pacing/timing, lighting, environment, and emotional tone. If dialogue/lyrics/subtitles matter, also call `transcribe_audio(audio_url=...)`. Use this analysis as the primary driver — do not redesign the scene or add characters/props/actions/text/locations that aren't in the reference. 3. **Establish character identity.** For each main character, generate a consistent reference image with `generate_image` (`model='t2i'`, `orientation='portrait'`), describing appearance in detail per `references/style-guide.md`. Pass `name` (e.g. "Mother ref", "Toddler ref") and `entity_type='characters'`. These refs are reused in every shot so identity never drifts. 4. **Generate each shot.** Break the video into shots (typically 5–8). For each shot call `generate_video` with `model='ref2v'`, `input_urls=[<character refs>]`, `orientation='portrait'`, and a prompt describing that shot's action, framing, camera and timing per the reference and style guide. Pass `name` (e.g. "Shot 1") and `entity_type='storyboard_videos'`. Keep the same characters, environment and lighting across all shots. 5. **Stitch the final cut.** Combine the shots in order with `mix_clips` (`orientation='portrait'`), matching the reference's total duration (e.g. 18s). Pass `name` (e.g. "Final cut") and `entity_type='final_videos'`. 6. **Optional audio.** If the user wants sound, generate a gentle, subtle instrumental with `generate_music` (no overpowering the visuals) and mix it in via `mix_clips`. This skill's default is no dialogue. 7. **Reply.** Summarize what was produced in natural language (don't paste large URL blocks), and offer to refine any shot, restyle a character, or re-mix. ## Notes - Use `write_todos` to track the multi-step pipeline, and `set_guidance` after the final cut to suggest next steps. - Report failures honestly; never fabricate URLs. Video generation does not support NSFW/adult content on any channel — politely decline such requests. - Intermediate products (character refs, shot clips) will appear on the canvas alongside the final cut; tell the user this up front. - If the user is unhappy with a shot, regenerate just that shot with the same character refs rather than redoing everything.