FILM & STORY · SKILL · POWERED BY MINIMAX-H3
AI Music Video Generator
for Storyboard-Led Tracks
Turn a track into a storyboard, connected scenes, and a cut you can refine. We read the beat, the energy curve and the mood — then your images and brief become a beat-locked video.
LIVE WORKSPACE · NO UPLOAD LEAVES THIS BROWSER UNTIL YOU GENERATE
Turn a track into a beat-locked cut
Analysis runs locally with the Web Audio API. Generation uses MiniMax-H3 (native audio + video).
We read rhythm, tempo, structure, mood and the energy curve — no server upload.
Their order is the shot order. Auto-compressed to JPG · 0.00 MB total.
reference_image items and the trimmed track becomes a reference_audio item in one POST /v2/video_generation call. H3 renders picture + sound together; cuts land on the detected beat grid.Upload a track and at least one image — shots are generated automatically on detected downbeats.
Local render: canvas + your original audio, no API call. Ideal for locking the cut before paying for an H3 render.
Uses POST /v2/video_generation · polls /v2/query/video_generation/{taskId} every 10s.
PRODUCT TEARDOWN
How the track becomes a cut
The original product states its job clearly: “At Buzzy, we treat the audio as the starting point for the visual work.” Its generator “analyzes a track’s rhythm, tempo, structure, and mood, then uses that reading to organize a concept, story, shots, pacing, motion, and transitions.” Here is what each of those claims maps to in this implementation.
| Capability | Original product wording | Implementation here |
|---|---|---|
| Rhythm & tempo | “analyzes the rhythm, tempo, structure, and mood” | 512-point FFT spectral-flux onsets → autocorrelation tempo → phase-scored beat grid; 4/4 bar markers |
| Song structure | “chorus, verse, change in energy, or instrumental passage” | 8s / 16-beat segments labelled Intro · Verse · Build · Chorus / Drop · Break · Outro from RMS slope |
| Mood & look | “name the feeling and the visual world” | Valence × arousal quadrant from spectral brightness, BPM, mean energy and dynamics → palette, camera language, per-shot mood |
| Cuts on the music | “cuts, motion, and transitions can follow the music” | Cut points snapped to detected downbeats; every shot carries energy, beat accents and a camera instruction |
| Best moment first | “start with the audio you want to see” | Sliding 15s window scored by energy + onset density + upward drive → auto-picked clip start |
| Reference consistency | “add a reference image … keep the same look, outfit, and style” | Up to 9 ordered images become H3 reference_image items with explicit identity / wardrobe continuity rules |
| Brief & direction | “mood, color, lighting, camera style, and overall direction” | Storytelling / Performance / Lyric / Visualizer / Social-first × 8 style presets × free-text brief |
| Export frame | “16:9, 9:16, and 1:1” | 7 H3 ratios incl. 21:9; local WebM preview + storyboard JSON; MP4 from the H3 task |
WHAT YOU GET
Built around the musical shape
Automatic beat detection
BPM, beat grid, downbeats and onset density from spectral flux — not a fixed interval.
Energy & emotion curve
A readable RMS envelope with labelled song sections and a valence × arousal mood map.
Beat-snapped storyboard
Up to 9 ordered images become shots with their own energy, camera and mood instructions.
One H3 render
Images + trimmed audio + a structured storyboard prompt in a single native audio-video call.
HOW TO USE
Three steps from a song to a cut
STEP 01
Upload your song
Bring an MP3 or WAV. The generator analyses rhythm, tempo, structure and mood, then picks the strongest 15-second window.
STEP 02
Describe what the song should become
Choose a storytelling or performance direction, add ordered reference images, set a style, aspect ratio and cut density.
STEP 03
Generate, review, and adjust
Preview the beat-synced cut locally, edit any shot prompt, then render with MiniMax-H3 and download the MP4.
ENGINE SPECS
MiniMax-H3 parameter contract
| Parameter | Value |
|---|---|
| Model | MiniMax-H3 (native picture + audio output) |
| Create endpoint | POST /v2/video_generation |
| Query endpoint | GET /v2/query/video_generation/{taskId} |
| Regions | api.minimax.io · api.minimax.cn · api.minimaxi.com |
| Duration | Integer 4–15 seconds |
| Resolution | 768P · 2K |
| Aspect ratios | adaptive · 21:9 · 16:9 · 4:3 · 1:1 · 3:4 · 9:16 |
| Reference images | Up to 9, 256–5760 px, ratio 0.4–2.5, ≤30 MB each |
| Reference audio | Up to 3, MP3 / WAV, ≤15 MB — requires ≥1 image |
| Request body | ≤64 MB (base64 media), prompt ≤7000 chars |
| Poll interval | ≥10 seconds |
| Credential | Pay-as-you-go / Credit API key (not OAuth / Token Plan) |
WHY H3 FITS MUSIC VIDEO
Audio is an input, not an afterthought
Older image-to-video endpoints accept a still plus text and return silent frames. H3 accepts reference_audio alongside reference_image items, so the model sees the actual rhythm while it plans motion and generates its own soundtrack. That is why this page sends a trimmed WAV of exactly the selected clip window rather than a timing description.
H3 also allows a single structured storyboard prompt with timestamped shots mapped to ordered reference images — the same “master timeline + micro-timeline” pattern the official MiniMax prompting guide recommends.
FAQ
Questions
START WITH THE AUDIO
Let the song set the pace.
Upload a track, hear the beat grid, lock the cut, then render it with MiniMax-H3.