---
name: musicvideo
description: "Reference guide for an AI music video on a whole track starring the person's own recurring AI character — the source reference method: one film you have permission to use as the world, a linear three-act story, every face replaced (the person's character as the lead), MiniMax H3 Max reference-to-video in ≤15 s parts of 8 shots with transitions, light, acting and constant lip-sync. Points to musicvideo-* sub-skills with exact settings, prompt structures and times. Use when the user says music video, clip for a track, клип, муз клип, видео на трек, клип с персонажем. Two external routes: the measured H3 method and a separate draft option, plus the quality route — Seedance 2.5 reference-to-video; the source preferred Seedance for quality, but the side-by-side measurement was a product commercial, not every format."
---

These creative references describe source experiments with external tools. Use your own supported account and tools; this download installs no automation. Settings and timings are experimental observations, not current FlashBoards availability. Use only media and voices you have permission to use.

# musicvideo · reference workflow

**Work for the person's own character and track: their photo / face board / look sheet from their character card. The example run below belongs to a demo character.**

Slots in all musicvideo-* skills: `{character}` — the person's character (a story may have more than one person in it), `{face board}` — the 5-angle face board of `{character}`, `{look sheet}` — the costume/prop sheet, `{track}`, `{film}`, `{act}`. Values in [brackets] are examples, not defaults.

## 🏆 Source experiment pick — the method

0. **Cast first.** The person's `{character}` is the lead. If the chosen `{film}` has a second main role, ask the person who plays it: another of their characters, an invented face, or removed. **Never cast a demo character.**
1. **The track** → lyrics with word times (Scribe v2), style (fal audio-understanding), the beat grid measured from the audio, parts ≤ 15 s cut on a beat at a vocal pause → `musicvideo-track`.
2. **One film** whose story fits `{track}`; a **linear story in three acts** (joy → turn → catastrophe/ending); frames for every story beat from the film's HD master → `musicvideo-film-scenes`.
3. **Every face replaced**: the lead → `{character}` (head swap, the film's costume may stay), other main roles as cast in step 0, everyone else → invented faces; a new costume or prop is a `{look sheet}` first → `musicvideo-replace`.
4. **Video**: H3 Max reference-to-video, 16:9, 768P, enhance disabled; Image 1 = `{face board}` (5 angles), Image 2 = `{look sheet}`, then the frames; **8 shots a part**, cuts on beats, a different transition from the H01 glossary on every cut; light, action, acting, head movement and lip-sync written for every shot; all parts run in parallel → `musicvideo-shots`.
5. **Edit**: trim each part to its length, join, lay the original track; one final upscale → `musicvideo-edit`.

**Write every prompt fresh for this track and story; the example prompts are for reading only.**



## Two external routes: draft or quality

| Route | Model | When |
|---|---|---|
| **Separate draft option** — not the measured 768P method above | MiniMax H3 Max / H3 Max Turbo, 480P | drafts and many takes; not the measured music-video setup |
| 🏆 **Quality** | **Seedance 2.5** reference-to-video (`bytedance/seedance-2.5/reference-to-video`, `task: reference`): the same pictures as `image_urls` (@Image1…), the same audio as `audio_urls` (@Audio1…), 4–30 s a call, native synced sound, 480p / 720p / 1080p | the final cut, when the look matters |

**Source observation (9.10.2026): Seedance 2.5 was preferred for quality; H3 remains the draft route.** Measured side by side on a 15 s product commercial (`ad-spot`): Seedance 2.5 at 720p — cut deviations up to about 1 s in the listed measurements (versus up to 2 s on H3), finer liquids and detail, its own sound effects in sync — took 7 min against 9 s on H3 Max 480P. Not yet measured on every format; write the same shot-by-shot prompt in plain text and refer to the references as @Image1, @Audio1. Compare the routes before choosing a generation.

## Classify

| The job | Go to |
|---|---|
| Lyrics, style, tempo, where to cut the track | `musicvideo-track` |
| Choose the film, find the scenes and frames | `musicvideo-film-scenes` |
| Put the character into the frames, remove the actors | `musicvideo-replace` |
| Write the prompts and generate the parts | `musicvideo-shots` |
| Assemble, sound, upscale | `musicvideo-edit` |

## What we tried and dropped (all on the demo track)

| Approach | Why dropped |
|---|---|
| Choppy “fever dream”: 12 cuts a part, reference locations | no story |
| Story with invented locations and friends | the outfit jumps between characters, no film feel |
| Scenes from three different films | not one world |
| H3 Turbo extend-video, frame per shot | at 21 s, but only a still frame starts each shot |
| H3 Turbo extend-video, text only | faces drift, the film is lost — **not for music videos** |
| Selfie reel tied to the frames | camera copies the film framing; a frontal face reference freezes the head |

## Global rules

1. **One film, one linear story.** No jumps between stories; supporting characters only as background.
2. **No actor's face on screen**, not even an extra's — check with crops; GPT keeps famous faces unless you crop and edit them.
3. **Props and costumes as pictures**, never as words (a look sheet → the frame edit → Image 2).
4. **Blind prompts**: picture roles only, never describe the person's looks or the content of the frame. Exception (owner, 9.10.2026): the expression of a reference frame — gaze, brows, mouth — may be named.
5. **A 5-angle face board as Image 1**, aspect ratio 0.4–2.5 (lay it out in two rows).
6. **Never mention phones** — the model draws them.
7. The source used cloud tools (fal, Magnific) for heavy compute and local ffmpeg to trim and join. Use your own supported environment; no Mac or author-specific runner is required.
8. **Never reuse the example run's files, prompts, faces, frames or story** for a person.

## Example run (reference only — never reuse its files, prompts, faces or frames for a person)

9.10.2026, demo character **Mara Vale**, clip **«The Welder»**, track **«3upd (Cover)»**, 88.8 s → 7 parts (0–11.81, 11.81–26.30, 26.30–41.24, 41.24–52.56, 52.56–64.78, 64.78–79.27, 79.27–88.80). Film *Titanic* (1997) in three acts: boarding, sunset at the bow, the party below deck → the boiler room (sabotage) → the flood, the corridors, the lifeboat as the ship sinks. Mara in a welder's look sheet; Cal and Jack survived replacement and were fixed by crop-and-paste / removal.
Seven parts in parallel took ≈ 35 s; the final clip was upscaled once.
Article with every prompt and frame of that run: https://flashboards.yaroflasher.com/learn/formats-and-exhibition/ai-music-video/

No example track, private film corpus or local experiment files are delivered or required. Choose your own authorized film and track, and prepare your own scene notes and outputs.