RunningHub MiniMax H3 Ref2VA Audio Reference (Legacy)

Here's the thing most video models just can't do: this node hands MiniMax H3 a piece of audio and H3 treats it as a reference to honor, not a file to ignore. Because H3 generates audio jointly with video in a shared context, an audio reference can carry a voice to match, a soundscape to continue, or a style of audio to replicate. It's the difference between "generate some sound" and "generate sound like this."

Inputs

• audio - required, an AUDIO input. Load an audio clip with any ComfyUI audio loader. This is your reference: the voice, the ambience, the instrument, whatever sound identity you want the generation to lock onto.
• references - optional, a MINIMAX_H3_REFERENCES object. Same chaining trick as the image reference node: connect a previous reference node's output and this appends the audio to the running stack.

Output: references - the combined MINIMAX_H3_REFERENCES list.

What you can actually do with it

The killer use case is character and voice consistency: build a reference stack of a character image plus a voice clip, prompt an action, and H3 returns video where the character looks right and sounds right. The model's launch reception specifically highlighted this unified-context behavior - audio matched to the picture, generated in the same pass. If you're coming from a text-to-video tool that bolts on a TTS voice afterward, this is a different, much smoother workflow.

Two practical notes. The reference order in the stack matters - the target and encode nodes fingerprint it, so keep the chain order stable between runs. And the audio length is a real input to the target: the Ref2VA target node will happily derive the clip duration from a single audio reference (set duration_seconds to 0 there). Use that as a shortcut to make the generated clip exactly as long as your reference audio.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/RH-RunningHub/ComfyUI-RH-MiniMax-H3.git
pip install -r ComfyUI-RH-MiniMax-H3/requirements.txt

Restart, and pull the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/. Ref2VA also needs system ffmpeg and ffprobe on PATH for audio/video reference prep - that's in the README and it's a genuine requirement. Legacy node; RHMiniMaxH3RefGen wraps this workflow today. License as pack-wide: H3's Community License excludes the US, EU, UK and South Korea from local-weights use.