JZL - 🎧 音频调度

"JZL - 🎧 音频调度" (Audio Dispatcher) is the pack's dynamic audio router. In the JZL short-drama pipeline, the Script Processor writes per-segment scheduling instructions, and this node reads the audio part of those instructions and decides which of your loaded reference audios each segment actually gets. H3 is one of the first open video models with native, jointly-generated audio - which means reference audio slots are a real thing worth routing carefully, not an afterthought bolted on after rendering.

How it works

The single required input is va_instruction (a wildcard *, force-input) - the audio scheduling string from the "剧本与镜头处理器" (Script Processor) or the segmented text chain. Inside, the node parses it into an ordered list of slot names. Then it looks at the optional inputs you've connected - the _FlexibleInputType means it accepts as many AUDIO-typed inputs as you hang off it, dynamically - and matches each slot name to a connected audio by name. The matched audios come out in slot order on ref_audio_0, ref_audio_1, ref_audio_2, which feed straight into the ref2va encoding path (ref_audio_N inputs on a "MiniMax H3 参考编码" node, or the ref bus).

The key behavior: output order follows the instruction's slot order, not the order you connected your audio inputs. And if there's no scheduling instruction at all (empty slots), it deliberately outputs None rather than dumping everything in - the pack calls this out as a feature so a missing instruction doesn't silently inject the wrong soundtrack.

The inputs that matter

• va_instruction - the scheduler string. This is the boss.
• The dynamic AUDIO inputs - these are your candidate reference audios. Up to three can be selected and emitted (ref_audio_0/1/2).

That's the whole surface. It's a routing node: no model, no audio processing, no GPU.

How to install it

Same pack as everything else - ComfyUI Manager, search "ComfyUI-JZL-MiniMax-H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/wjluoxiao/ComfyUI-JZL-MiniMax-H3

Restart ComfyUI. Zero extra dependencies for this node beyond the pack itself.

Common issues

Name matching is the trap. The slot names in va_instruction have to line up with the names of the connected audio inputs - the matching tolerates some fuzz, but if a slot names something you never connected, that output stays None and your clip gets no reference audio for that segment. If you find segments randomly going silent, print the instruction string and check the names. And remember the output is capped at three reference audios regardless of how many you connect; that's a H3 ref2va limit (the encode node itself only takes ref_audio_0/1/2), not a bug in the dispatcher.