MiniMax H3 VAE Loader

MiniMax H3 generates video and audio together, which means it needs two autoencoders - one for frames, one for sound - and this loader is how the workflow references them. It's the quietest node in the pack: no model math, no loading bar, no VRAM hit at all. It's a bookmark. You point it at the VAE files on disk, it hands the workflow an av_encoder reference, and the actual decode happens lazily later in the MiniMax H3 Decode AV node.

How it works

The display name says "Loader," but the source is explicit: it just builds a reference object holding the file paths. The video VAE and audio VAE weights are loaded by Decode AV on first use, which keeps your VRAM for the sampler where it belongs. Think of it as the middle man between your models/vae/ folder and the nodes that actually touch pixels and waveforms.

The inputs that matter

• vae_name - required, the H3 video VAE, a .safetensors from ComfyUI/models/vae/. This is the autoencoder that turns the video latent into frames, and the conditioning encoder that turns reference images into latents also runs through it.
• audio_vae_name - optional, default "none". Select the H3 audio VAE here if you want generated sound (and with H3's whole selling point being native audio, you usually do). Set it to "none" and Decode AV will hand back a silent audio track instead of failing.

The single output, av_encoder, is the type this pack calls MINIMAX_H3_AV_ENCODER. Wire it into two places: the MiniMax H3 Conditioning node's av_encoder input (so it can encode your reference images into the right space) and the MiniMax H3 Decode AV node's av_encoder input (so it can decode the joint latent). Both usually come from this one node, and it's fine to split the output to both sockets.

Installing it

The pack install is the same as its siblings - ComfyUI Manager, search "MiniMax H3", or:

cd ComfyUI/custom_nodes
git clone https://github.com/xiaolibai-sys/ComfyUI-MiniMaxH3
pip install -r requirements.txt

then restart. Dependencies are light (torch, safetensors, numpy). The actual work is dropping the VAE files into ComfyUI/models/vae/ before ComfyUI starts, so they appear in the dropdowns. The video VAE and audio VAE are separate downloads from the model's HuggingFace repo - grab both or you'll be missing either frames or sound.

Common issues

• No audio out. Nine times out of ten the audio_vae_name is sitting on "none". Pick the audio VAE file.
• File not in dropdown. The folder list is captured at startup. Put the VAEs in models/vae/ and refresh ComfyUI.
• Reference encoding complains. If Conditioning can't encode your images, check that the video VAE you pointed at is the H3 v2 VAE - the pack's tests align specifically against the H3 v2 encode, so an older H1/H2 VAE won't behave.

The one thing worth remembering: this node is deliberately boring. If you see it "doing" something, that's a feature - everything heavy is deferred to decode time.