RunningHub MiniMax H3 Decode Video + Audio (Legacy)

You made it through the sampler - now where's the video? This is the node that answers that. RHMiniMaxH3DecodeAV is the endpoint of every MiniMax H3 generation: it takes the denoised AV latent and runs it through H3's two VAEs to produce actual frames and an actual audio waveform. Until this node runs, everything you've built is just a big tensor, and plenty of beginners stare at a sampled latent and wonder if they broke something. You didn't. You just haven't decoded yet.

Inputs

• h3_vae_bundle - the dual VAE bundle. Video goes through the 24-channel video VAE, audio through the 32-channel audio VAE, sequentially - and the bundle has to match the latent you're decoding. The node checks fingerprints and throws a clear error if you've mixed a latent sampled with one checkpoint against another bundle's VAEs. That's a feature: it saves you from silent garbage.
• sampled_av_latent - the output of the Dual Sigma Sampler. If you plug an empty (unsampled) latent in here, it refuses with "input is still an empty latent - connect the Dual Sigma Sampler first," which is the friendliest error in the pack.

Outputs:

• frames - IMAGE, the video batch. Wire to a VHS-style video output node or a preview.
• audio - AUDIO, the waveform. Combine it with the frames into an mp4/webm, or listen to it alone.

How it works

The decode runs the two VAEs one after another - video first, then audio - and the source notes that this is a hard residency boundary: the DiT and Qwen encoder leases are offloaded before video decode begins, and completed temporal chunks accumulate on CPU. What that means for you: decoding is the moment your VRAM gets to breathe, and it's also where the pack writes its telemetry sidecar metadata. Nothing you need to do, but it's why a long clip can look like it's hanging right at the end - it's decoding in chunks, not frozen.

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 have the ~95 GiB INT8 ConvRot bundle in ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (HF Gluttony10/MiniMax-H3-INT8-CONVROT, ModelScope in China). ComfyUI 0.27+ and CUDA PyTorch are the floor. It's a legacy node - the modern RHMiniMaxH3VideoGen bundles generation and decode - but old workflows with this node load and run fine. And the pack-wide reminder: H3's Community License excludes the US, EU, UK and South Korea from local-weights use.