Lip-Sync MV Maker · সর্বজনীন Open Skill

Lip-Sync MV Maker

Lip-Sync MV Maker Digen-এ একটি সর্বজনীন Open Skill। খুলে দেখুন এই ওয়ার্কফ্লো কী তৈরি করে।

@waytitiai

Music

3 1 3

https://d7qlwi7tl2u27.cloudfront.net/skill-covers/prod/4210253/2026-10-21/93kkrhnv.jpg

---

allowed-tools:
  - probe_media
  - extract_audio
  - extract_frame
  - transcribe_audio
  - recognize_image
  - recognize_video
  - generate_video
  - generate_music
  - generate_tts
  - mix_clips
  - caption_video
  - fetch_video_info
  - download_video
  - create_document
  - update_document
  - get_document
  - memory_query
  - memory_store
---

# Lip-Sync MV Maker

Turns a still photo + a song into a music video where the subject performs (singing, lip-sync, or dancing) in time with the audio. Also covers MV reimaginings and parody MVs with custom lyrics.

**Operating principle — FREEDOM**: the creative direction is the user's. Pick sensible defaults, offer concrete options instead of long questionnaires, and never lock the project into one genre/mood. Never spend credits on things the user didn't ask for.

## Credit safety (top priority — never drain the user's balance)
1. **Free tools first.** Analysis, timing and assembly are free/cheap: `probe_media`, `transcribe_audio`, `recognize_image`, `recognize_video`, `extract_audio`, `extract_frame`, `extract_clip`, `mix_clips`, `fetch_video_info`. Exhaust these before any generation.
2. **Plan + cost estimate before the first generation.** The plan states the exact number of video generations, the audio generations, and the total seconds. Nothing is generated before the user has seen that number.
3. **One generation per planned shot.** No candidate batches, no "let me try 3 versions", no silent test renders. The planned count is a hard ceiling — if the project needs more, ask first and offer a cheaper downgrade (shorter shots, fewer shots, reuse an existing shot) before spending.
4. **Respect an explicit budget.** If the user states a cap ("max 4 generations", "keep it cheap"), stay inside it and say when the cap is reached.
5. **Reuse assets.** One source photo for every shot; one TTS/music take per track; never regenerate an asset that already exists and still fits.
6. **Cheap settings by default.** Video at `720P`; raise resolution only on request. Do not add an upscale, enhancement, or extra render pass unless the user asks.
7. **Retries are capped.** Maximum one retry per shot, and only for a shot the user wants fixed — never retry shots that are already acceptable.
8. **Report actual usage.** After delivery, state how many video/audio generations were spent versus the estimate.

**Cost discipline**: plan first, generate once. Never generate a "test batch" of variations. Keep the number of video generations to the number of shots actually planned, confirm with the user before exceeding the planned shot count, and reuse the same source photo for every shot.

## Engine default (keeps the face and the lips intact)
- **Default = `ai2v` (photo + audio) with restrained motion and a steady camera angle.** This is the best combination of face retention and clear lip-sync.
- **Escalate to `ref2v` only when it is genuinely needed**: choreography must follow a reference MV, or the scene needs wide/full-body movement. Expect slightly weaker lip-sync fidelity there — say so, and keep `ref2v` shots short.
- Shot length default **10s** (max 14s on `ai2v`). A 30s track → 3 shots of 10s.
- Keep the singing face large in frame (medium close-up). Wide shots and fast motion degrade mouth readability — flag it before generating, not after.
- Full multi-shot protocol, safe vs risky motion lists, emotion and weather prompt clauses: see `references/mv-recipes.md` → "Multi-shot scene direction".

## When to use
- Photo + song → singing / lip-sync / dancing MV.
- "Reimagine this MV", "parody MV with my own lyrics", "cover version of this song".
- User uploaded a reference MV video and wants their own character to perform the same choreography/mood.
- Only one of photo/audio is missing → offer to generate the missing one (music via `generate_music`, vocal via `generate_tts`).

## Steps

1. **Clarify (max 1 round, offer defaults)**
   Collect: reference photo(s), audio (uploaded song or link),