APNext H3 Music Video (Minimal)

Here's the whole pitch: drop in a song, paste the lyrics, pick a cinematic look, and the node writes the entire music video. Concept, performer, staging, shot list - the model invents it all. You get one run with three sliders and a clean conscience.

The "Minimal" is doing real work there. This is a one-box wrapper around the pack's full H3 Music Video Writer, which is a deep node with cast sockets, wardrobe locks, location locks and a masked-audio path. The minimal version calls that writer with opinionated defaults so you don't have to. It's a MiniMax-H3 prompt writer underneath, not a renderer - it outputs the scene prompts and audio slices; a MiniMax H3 video node renders them, and H3 Scenes Join stitches the clips back into one video with the original song. The pack ships h3_music_video.json and h3_music_video_minimal.json showing the whole chain end to end.

How it works

Your song is cut into consecutive pieces, each no longer than H3 renders in one clip (5–15 s), and the cut points are chosen on the music - onsets, energy steps, section changes, and right before a lyric line when the lyrics are timed. Each piece's length is snapped to H3's frame grid, so every rendered clip is exactly as long as its audio slice and the stitched video never drifts. Each piece becomes one scene: it's tagged <Audio 1> and reused 1:1 as that clip's soundtrack, and the scene's imagery is staged from that piece's lyric lines.

The three sliders map onto the full writer's controls. performance (default 80) decides how much the singer is on camera: 0–33 is Narrative (story visuals, nobody sings on camera), 34–66 Mixed, 67–100 Performance, where the singer lip-syncs the actual lyric lines on camera. pace (default 30) sets how fast it cuts - the code maps it straight onto max segment length, from ~15 s per clip at 0 down to ~6 s at 100. wildness (default 45) runs the pack's usual scale: grounded performance video at 0, fully surreal above 40, where the seed decides which weird events creep in.

The inputs that matter

• audio - the song (load it with Load Audio). This one's required, obviously.
• lyrics - one line per line. Timestamps make the sync exact: [0:15] line, 0:15 line, or LRC [00:15.20] line. [Chorus] tags are kept; untimed lines are spread evenly. Empty = instrumental video.
• visual_style - the whole video's look. The curated cinematic styles each fix camera, lenses and colour; Auto lets the model pick one to fit the song.
• performance, pace, wildness - the three knobs above.

Reference images (image_1..9) are the one thing worth attaching: they fix the performer's face, and they pass straight through to the video node. Without them, the model invents a performer from the lyrics.

The writing happens in your local Claude Code CLI (see the pack's Claude Code nodes - subscription seat, no API key, no model files), or on any local/API model if you connect an APNext H3 LLM Backend to the llm socket. The model dropdown lists the Claude Code aliases (sonnet / opus / haiku / default) plus Codex and anything your local servers were serving when the page loaded.

Outputs and the honest catches

Wire these to the video node: scenes (one prompt per piece), lengths (H3 frame counts - no math node needed), audio_segments (the matching audio slice per piece, into ref_audio_1), and clip_starts if you're doing your own assembly. scenes_text is the whole thing as readable text for previewing.

Where people get caught out: a long song means lots of clips, and each one is a separate H3 render - a 4-minute track is a serious queue, not a quick test. The node is only as good as its lyrics sync: untimed lines get spread evenly, so a track without timestamps loses the lip-sync precision that's the whole fun of Performance mode. And it uses your Claude Code subscription seat, so a heavy session can hit the rolling 5-hour rate limit mid-run. If you outgrow the four knobs - you want a written concept, a specific cast, wardrobe or location locks, or the masked-audio path for structural lip-sync - graduate to the full H3 Music Video Writer. This one is for the 80% case: song in, video out.