Nodes/CRT-Nodes/MiniMax H3 VIDEO VAE (CRT AutoDL)
ComfyUI Node

MiniMax H3 VIDEO VAE (CRT AutoDL)

H3's video VAE, no inputs, one download, done

By PGCRT·Created 2 years ago·Updated 6 days ago· 138
MiniMax H3 VIDEO VAE (CRT AutoDL)
    • VAE

    Every video model needs a VAE, and MiniMax H3 VIDEO VAE (CRT AutoDL) is MiniMax H3's. It's the simplest possible node - no inputs, one VAE output - and the only thing it does beyond loading is save you the file hunt: on first run it downloads minimax_h3_video_vae_fp16.safetensors from Comfy-Org's MiniMax-H3 repo into ComfyUI/models/vae/ and loads it.

    That's the entire CRT AutoDL promise in one node, and it's genuinely useful precisely because it's boring. H3's checkpoints are huge, the repackaging story across Comfy-Org and Kijai repos is fiddly, and "which VAE do I even need" is a question beginners shouldn't have to answer. This node answers it: the fp16 video VAE, the one ComfyUI's H3 support expects.

    Where it fits

    H3 is a two-VAE model - frames go through this video VAE, the jointly-generated audio goes through MiniMax H3 AUDIO VAE. A complete H3 graph has both: the video VAE decodes the frame latent into images, the audio VAE decodes the sound latent into a waveform. Wire this node's VAE output into the decode step on the video side of your sampler, use the sibling audio node on the audio side, and you're set.

    Because it's an AutoDL node with zero inputs, there's genuinely nothing to configure. Add it, let it download once, and it behaves like any other VAE loader in your graph. If you've already got the H3 model and CLIP loaded through their CRT nodes, this is the last piece of the pipeline.

    Installing it

    Same as every CRT node - install the pack once:

    cd ComfyUI/custom_nodes
    git clone https://github.com/PGCRT/CRT-Nodes.git
    pip install -r requirements.txt
    

    or ComfyUI Manager → search CRT-Nodes, then restart. The fp16 VAE is a couple of GB, so the first-run download is the only real wait.

    Common issues

    • Wrong VAE for the job - this decodes H3 video latents. Feed it another model's latents and you'll get noise or an error. Pair it with the H3 model, CLIP, and audio VAE from the same stack.
    • License - H3 weights are under the MiniMax H3 Community License, whose applicable territory excludes the US, EU, UK, and Korea. That restriction covers this VAE too.
    • Red node after an update - right-click → Fix node (recreate).
    CategoryCRT/AutoDL/MINIMAXH3

    Inputs (0)

    No inputs

    Outputs (1)

    NameTypeDescription
    VAEVAE