RunningHub MiniMax H3 Ref2VA Dual VAE Loader (Direct) (Legacy)

Every MiniMax H3 workflow needs its two VAEs - a 24-channel video VAE and a 32-channel audio VAE - loaded as one bundle. RHMiniMaxH3Ref2VAVAELoader was the Ref2VA-specific, explicit-root way to do that, back when the pack's loaders took a model_root and let you point at exact files. It's deprecated now, but if an old graph still carries it, here's what it's doing and why it got retired.

Inputs

Three required, all self-explanatory in the "Direct" style:

• model_root - the weights root. The dedicated models/MiniMax-H3-INT8-CONVROT (or legacy models/MiniMax-H3), or an official release root under models/diffusers. All three components must come from the same root.
• video_vae_path - MiniMax-H3-video_vae.safetensors, the 24-channel video VAE. The merged official artifact name, with the sharded raw-package logical name (MiniMax-H3-video_vae) also accepted.
• audio_vae_path - MiniMax-H3-audio_vae.safetensors, the 32-channel audio VAE, same naming story.

One output: h3_vae_bundle, the typed handle that carries both VAEs together.

The quirk that retired it

Like the text encoder, the video and audio VAEs are byte-identical across the FL2VA and Ref2VA partitions. So this node - which always resolves from the Ref2VA partition - was a partition-locked duplicate of a loader with no actual partition difference. The pack merged the pair into RHMiniMaxH3VAELoader and dropped the root/partition widgets entirely. The modern node still asks for the two VAE filenames, because those genuinely differ per file, but "which partition" stopped being a question anyone needed answered.

The honest take

Skip it for anything new. It's hidden from search and only registered so old saved workflows keep executing. If you inherited a graph using it and it's misbehaving, the migration path is the pack's migrator:

python3 tools/migrate_workflow.py old_workflow.json --in-place

That handles the older node-ID scheme, the old single vae_path that got split into video_vae_path + audio_vae_path, and widget-count drift. If the migrator doesn't cover your case, a manual swap to RHMiniMaxH3VAELoader with the same two filenames is a drop-in change - the output type is the same bundle handle.

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, then the ~95 GiB INT8 ConvRot bundle into models/MiniMax-H3-INT8-CONVROT/ - both VAE files ship inside it. And if a workflow errors with a missing video_vae, double-check the bundle's actual location before suspecting the node; the path-scan is strict.