APNext H3 Music Video Chain Render (carry frames between scenes)

Here's the hole this node fills. When the APNext H3 Music Video Writer emits a scenes list, ComfyUI's list processing renders every scene as an independent clip - scene 3 literally cannot see scene 2's frames, because both are items of the same list. So the person in the music video changes clothes between cuts, the room rearranges, the motion restarts. Real continuity needs the previous render, not the previous prompt.

H3MusicVideoChainRender renders the writer's scenes one after another inside one execution, carrying the last context_frames of each delivered clip into the head of the next one where the take should continue. It's a dagthomas/comfyui_dagthomas node, and it does in one node what the ComfyUI-MiniMaxH3-Contex-Loop pack does with a loop - the same continuity trick, but built around the masked-song workflow the pack's own examples use.

How it works

Per scene, the node: builds the Ref2VA conditioning (up to four ref_image pictures, the same for every scene), injects the song piece into the audio latent (the "masked audio" path - the song is written into the H3 audio latent and protected from denoising, which is what gives structural lip-sync), pins the previous clip's tail to the head of the new latent when the take continues, samples with your sampler / sigmas (scene k uses seed + k), decodes, trims the pinned head and grid padding off, and saves the clip with its audio piece.

Which boundaries continue is the Cut Plan's call, if you connect one: a cut placed on a drop, section start, stop, or one of your taps stays a hard cut - that's where the sync lives - while a cut on a mere onset, downbeat, or lyric line turns into a continuing take. Or force it with continuity (flow everywhere / cut everywhere).

The continuity mechanism prefers MiniMaxH3SongMaskedAVContext from ComfyUI-H3-Motion-Context-MultiRef (the pack that also powers the masked-audio example workflows), falling back to ComfyUI core's MiniMaxH3AddGuide - picture only, audio generated - with a clear console warning.

The inputs that matter

• model / clip / vae / audio_vae - from ComfyUI's MiniMax H3 loader nodes.
• scenes, lengths, audio_segments, clip_starts - the writer's list outputs; these are the four wires that make it a music video chain.
• master_audio - the whole song (Load Audio). Without it, H3 generates audio per scene.
• context_frames - how much of the previous clip's tail is pinned (default 22, ≈0.9 s of motion; snapped to H3's valid runs 5/22/39/56…; 39 also aligns the audio clock).
• carry - "previous latent" copies the sampled tail straight into the new latent (no decode/re-encode, nothing lost per link - the one you want); "previous frames" decodes and re-encodes.

Outputs: file_paths (list of saved clips), report, and audio (the joined scene audio, handy for wiring downstream).

Install

Pack plus one friend:

cd ComfyUI/custom_nodes
git clone https://github.com/dagthomas/comfyui_dagthomas
cd comfyui_dagthomas
pip install -r requirements.txt

Then install ComfyUI-H3-Motion-Context-MultiRef (Manager or git clone) for the full masked picture+sound continuity; without it the node still runs on the core guide and tells you so in the console. The H3 model itself loads through ComfyUI's own MiniMax H3 nodes - and heads-up, its weights are the big download (~40 GB+) with a community license that excludes the US, EU, UK and Korea, so check before you go all-in.

Troubleshooting

The common failure is a missing pack: connect master_audio, get a warning that continuity uses the core guide - install the Motion-Context pack and re-run. A scene piece too long to fit the pinned prefix under H3's 362-frame range opens on a hard cut with a warning instead of silently breaking, which is the kind of behavior you want. And when one scene in a finished run is wrong, don't re-run the whole chain - that's what H3 Scene Retake is for.