Nodes/MiniMax-H3 Tools/MiniMax H3 NF4 Audio VAE Loader / NF4 音频 VAE 加载
ComfyUI Node

MiniMax H3 NF4 Audio VAE Loader / NF4 音频 VAE 加载

Load MiniMax H3's NF4 audio VAE — the tiny 0.28 GB file that carries the soundtrack

By Rinne414·Created about a month ago·Updated about a month ago· 10
MiniMax H3 NF4 Audio VAE Loader / NF4 音频 VAE 加载
    • VAE
    vae_name

    The audio VAE is the piece that makes MiniMax H3 special - synchronized stereo sound generated in the same sampling pass as the picture, not bolted on afterwards. In the DiffSynth NF4 weight set it's a 0.28 GB file, tiny next to the 17 GB model and 15 GB encoder. But of the four NF4 files, this one needs the most massage before ComfyUI can use it, and this loader is where that massage happens.

    What it does

    The vae_name combo (populated from ComfyUI/models/vae), one VAE output. Drop the file in:

    | File | Folder | Size | |---|---|---| | audio_vae_nf4.safetensors | ComfyUI/models/vae | 0.28 GB |

    Behind that simple interface, the loader handles three quirks of the DiffSynth checkpoint:

    • Weight norm folding. DiffSynth stores the audio VAE's layers as weight-normalized pairs (weight_g / weight_v) instead of plain weights. The loader folds them into ordinary weights.
    • Latent statistics. It injects the audio VAE's published latents_mean / latents_std - the audio decoder needs those to normalize correctly, and they're not in the file.
    • NF4 attention. It restores the quantized Linear state and swaps the causal attention modules for NF4 QKV versions.

    Then it validates against ComfyUI's MiniMax H3 audio VAE structure and refuses to proceed on a mismatch. So when it works, you can trust it; when the file is wrong, it tells you instead of decoding garbage.

    Stack context

    This is the fourth of the four DiffSynth NF4 loaders - model, text encoder, video VAE, and this audio VAE. All four replace their stock loaders in place; keep the rest of the H3 workflow unchanged. Requires BitsAndBytes and CUDA:

    python -m pip install "bitsandbytes>=0.50.0"
    

    Same honest performance note as the others: NF4 saves disk (the whole stack is 34.4 GB vs 53.9 GB for the int8 route) but dequantizes for computation, so it measured slower on the pack's one RTX 3090 test rig. Size your stack by footprint, not speed.

    Install

    Part of the MiniMax-H3 Tools pack (ComfyUI 0.30.0+). Search "MiniMax H3 Tools" in ComfyUI Manager, or:

    cd ComfyUI/custom_nodes
    git clone https://github.com/Rinne414/ComfyUI-MiniMaxH3-Tools.git
    

    Restart ComfyUI, find MiniMax H3 NF4 Audio VAE Loader under MiniMax H3. Fails explicitly on FP4 files, missing quant sidecars, non-H3 files, and incomplete downloads (.incomplete / .aria2log artifacts are not model files). And remember the audio VAE carries the same MiniMax H3 Community License as the rest of the weights - that licence excludes the US, EU, UK and Korea from running the local files.

    Categorymodel/loaders/minimax

    Inputs (1)

    NameTypeDefaultDescription
    vae_nameCOMBO0 options:

    Outputs (1)

    NameTypeDescription
    VAEVAE