APNext H3 Music Video Writer

The moment you realise it: APNext H3 Music Video Writer turns a song into a music video. Not a mood video with a soundtrack slapped on - a video where every clip is cut on the music, every rendered piece is exactly as long as its audio slice, the performer lip-syncs the actual lyric lines, and the whole thing stitches back together against the original track. It's the most ambitious node in the H3 family, and it mostly works because it's ruthless about one thing: audio lengths and clip lengths must match exactly.

Here's the mechanism. H3 renders clips up to about 15 seconds, so the node cuts the song into consecutive pieces inside min_segment_seconds–max_segment_seconds. The cut points come from the music itself - spectral-flux onsets, energy steps, section changes, and (when you supply timed lyrics) right before a lyric line. Then every piece length is snapped to H3's frame grid (5 + 17k frames at 24fps) so each rendered clip is exactly as long as its audio slice. That's the detail that makes the final stitched video never drift out of sync - which is the failure mode that kills every hand-rolled attempt at this.

Then it writes one scene per piece, each a four-section H3 prompt where the piece is <Audio 1>, reused 1:1 as the clip's soundtrack. In Performance mode the singer lip-syncs the piece's lyric lines on camera (<Subject 1> sings <d>[English] exact line</d> in sync with <Audio 1>); Narrative answers the lyric with pictures; Mixed alternates. Quiet pieces get long intimate shots; loud or peak pieces get more cuts and the chorus look. Long songs are written in chunks of six scenes that continue one session, so the synopsis, wardrobe and location locks stay consistent.

The inputs that matter

• audio - the song (from Load Audio). Everything flows from this.
• direction - the concept: who performs, where, the look, the story arc, what the chorus looks like vs the verses.
• lyrics - one line per line. Timestamps make the sync exact: [0:15] line, 0:15 line, or LRC [00:15.20] line; [Chorus]-style tags are kept. Untimed lines get spread evenly (approximate). Empty = instrumental.
• performance_mode - Performance / Narrative / Mixed.
• segment_mode - Auto cuts on the music inside the allowed range; Fixed takes the longest allowed piece each time; Lyric lines tries hardest to cut right before a line.
• cast_1..4 / extra_cast - the performer(s). An H3 Characters node in ✏️ custom mode with a wardrobe is made for this - a described performer with a locked outfit.
• wardrobe / locations / enforce_wardrobe - the same locks as the Crossover Writer (cast-carried wardrobe is merged in).
• The usual visual_style, dialogue_language (the lyric language), wildness, and the Claude Code block, plus image_1..9 reference pictures.

The outputs are matched lists, which is the whole architecture: scenes → the video node's prompt, lengths (frame counts) → length, audio_segments → ref_audio_1, durations → your frame math. Plus segment_table - the cut list, 01 0:00.00 – 0:15.08 (15.08s, 362 frames) energy: peak lyrics: … - and scenes_text, synopsis, cast, scene_count, song_seconds, session_id, info.

The finishing move

Each piece renders as its own clip, so the last step is H3 Scenes Join with replace_audio set to the original song - the joined per-scene audio gets swapped for the real track, and one Create Video → Save Video writes a single music video that plays in sync. The pack ships examples/h3/h3_music_video.json end to end.

Installing

Part of dagthomas/comfyui_dagthomas: ComfyUI Manager → comfyui_dagthomas, or clone into custom_nodes + pip install -r requirements.txt, restart. Under APNext → H3.

Claude Code must be installed and logged in - the default engine, subscription seat and all - unless you connect an H3 LLM Backend to the llm socket and run it on Ollama instead. Budget for the slow end: a 3-minute song is a dozen-plus pieces, each a full Claude turn, plus the audio analysis. Where people get burned: forgetting timed lyrics and wondering why the "sync" is approximate, and skipping replace_audio and wondering why the final video's sound doesn't match the song. Both are one input away from right.