RunningHub MiniMax H3 Empty AV Latent (Legacy)

Think of this as H3's version of Empty Latent Image - the node that mints the blank latent everything else plugs into. The twist: it's an AV latent, so it's really two blanks in one object: a zeroed 24-channel video latent and a zeroed 32-channel audio latent, sized from a target you hand it.

The only input is target (a MINIMAX_H3_TARGET from the FL2VA or Ref2VA target nodes), and the output is av_latent - a shape carrier, not noise. The comment in the source is the whole story: "the sampler replaces these zeros with the two independently-seeded source-compatible noise streams." So this node's only real job is to be the right-shaped vessel for the Dual Sigma Sampler, which is why its input is a target and nothing else.

Where it sits in the graph

Target node → Empty AV Latent → Dual Sigma Sampler → Decode AV → frames + audio

That's the minimal T2VA/FL2VA/Ref2VA spine. The target determines resolution, duration, frame count and latent shape; this node makes the empty latent; the sampler fills it; decode turns it into pixels and sound.

The V2A wrinkle

There's a second use hiding in the tooltips of sibling nodes. For video-to-audio, you don't want a blank vessel - you want an AV latent whose video half is a clean encoding of your footage, so the sampler only denoises audio. That's RHMiniMaxH3EncodeVideoAVLatent's job: it takes an existing AV latent (often an Empty one) plus your frames and returns an av_latent with the video half filled in. Keep that in mind if you chain Empty → Encode Video → Sampler and set denoise_video=False.

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 ComfyUI. This node itself needs no models - but the whole pack does: the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (HF Gluttony10/MiniMax-H3-INT8-CONVROT, ModelScope in China). It's a legacy node, so new workflows may prefer the all-in-one RHMiniMaxH3VideoGen, but the Empty → Sampler → Decode spine is exactly what the modern node does internally, and understanding it makes everything else legible. License as pack-wide: H3's Community License excludes the US, EU, UK and South Korea from local-weights use.