MiniMax H3 Silent Audio Master (Internal)

H3's headline trick is that it generates audio with the picture. So asking for silence is more deliberate than it looks - and in this pack it's a design decision you should know about, because it's the difference between two very different kinds of silence.

You could get "silence" by just not saving the audio. You could get a kind of false silence by turning off audio handling and letting H3 fill in whatever it thinks belongs there - footsteps, room tone, a hum. What you can't easily get is a track that's guaranteed zero for exactly the right duration next to video segments that may or may not have their own sound. That last one is what this node makes.

It's the counterpart to Locked Audio Master: where that one restores your original recording at the full output duration, this one generates a duration-exact silent track at the same span, so the video you save has a proper audio stream instead of a missing one.

How it works

The logic is small and mostly about getting the length right, so the file doesn't come out short.

Duration resolution runs the same three-step ladder as its sibling: the plan's target_output_frames if set, otherwise the sum of segment lengths minus every overlap after the first, otherwise the first segment's length. Each of those corresponds to a real generation shape - auto-segmented, manually segmented, or a single segment - and the fallback chain is why the node works without you telling it which one you're doing.

Then it builds the track: a waveform of zeros, one channel, one sample-rate field at 44100, with a sample count of round(frame_count / 24 * 44100) and a floor of one sample so an empty waveform can never escape. That's it. No encoder, no model, no VAE - this node doesn't touch H3 at all, which is worth saying out loud because it means it can't fail for audio-model reasons.

The one thing it will refuse: a plan with no resolvable frame count. "The silent-audio interval has no target frames" means your plan is a stub, not that the node is broken.

Inputs and outputs

One required input: finite_plan (MINIMAX_H3_FINITE_SEGMENT_PLAN).

One output: Silent Audio (AUDIO), exactly as long as your video. In the expanded sampler graph it substitutes for the assembled audio at the very end, after the frames are final.

The important context is when this node is in the graph at all. When a plan disables video audio, the pack doesn't just stop asking H3 for sound - the changelog is explicit that doing so "fixes the AV audio stream and final master to silence instead of asking H3 to generate replacement sound." So H3 still conditions on a silent audio stream, and the final master is silent too. An explicitly uploaded locked audio asset takes priority over this path, so if you have both, locking wins.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/Songssx/ComfyUI-MiniMaxH3-TimelineDirector.git

Restart, search MiniMax H3. No dependencies to install - the project's pyproject.toml declares an empty list, and this node in particular only needs NumPy and PyTorch, both already in ComfyUI. It's an (Internal) dev-only node: enable dev mode in ComfyUI settings if you want it in the node menu. Prerequisites for the pack overall: recent ComfyUI with native MiniMax H3 nodes, the H3 Ref2VA model, CLIP, video VAE and audio VAE, Python 3.10+.

Common issues

Video saves with no audio track at all. You skipped this node and didn't wire the plan's audio anywhere. It's internal - plugin-built graphs insert it, hand-built ones don't.

"The silent-audio interval has no target frames." The plan has no target_output_frames and no usable segment lengths, which means you're feeding it something that isn't a real Segment Plan - commonly a Material Plan by mistake, since they're different types from the same planner.

You wanted generated sound and got nothing. Then video audio is disabled in your plan, or a locked asset took priority. Check the plan's audio settings; per-segment audio settings come from the planner, and this path is chosen for you based on them.

Sample rate mismatch complaints downstream. It's always 44100. If your other audio is at 48000, that's a rate conversion somewhere else in the chain, not a mis-set input here.