RunningHub MiniMax H3 FL2VA Dual VAE Loader (Direct) (Legacy)

This is the FL2VA-pinned version of the pack's dual VAE loader, and it's the most "why does this exist" node in the family. The VAEs themselves are shared across H3's partitions - there's one video VAE and one audio VAE in the bundle, used by FL2VA, Ref2VA and T2VA alike. The only thing "FL2VA" does here is tell the node which weights directory to prefer when it resolves the root. The node's behavior is otherwise identical to RHMiniMaxH3DirectVAELoader.

So, honest framing: if you're building anything new, don't seek this node out. It's a legacy compatibility node kept alive so old FL2VA graphs load without "Node type not found" errors. What's not legacy is the job it does - H3 genuinely needs two VAEs loaded as one bundle, and every generation path in this pack depends on that.

The two VAEs, quickly

• video_vae_path - 24-channel video VAE, default MiniMax-H3-video_vae.safetensors. Turns video latent back into frames (and frames into latent, for V2A).
• audio_vae_path - 32-channel audio VAE, default MiniMax-H3-audio_vae.safetensors. Turns the audio latent stream into the actual waveform.

Plus model_root for where they live. All defaults are correct; the only reason to touch them is a custom layout or a manual download. Output: one h3_vae_bundle handle feeding your encode and decode nodes.

The pattern worth internalizing

H3 keeps video and audio as separate latent streams for the entire sampling process and merges them only at decode. That's why this pack has "AV latent" as a first-class object instead of just a normal latent, and why the Dual Sigma Sampler has separate shift values for the video and audio streams. Grab this mental model and every other node in this pack makes sense - including why a loader for two VAEs is a thing at all.

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 get the ~95 GiB INT8 ConvRot bundle into ComfyUI/models/MiniMax-H3-INT8-CONVROT/ (hf download Gluttony10/MiniMax-H3-INT8-CONVROT --local-dir ./models/MiniMax-H3-INT8-CONVROT, ModelScope for China). ComfyUI 0.27+, CUDA PyTorch. New workflows: use RHMiniMaxH3VAELoader; keep this one for your old graphs. License as ever: MiniMax H3's Community License excludes the US, EU, UK and South Korea from running the local weights.