H3 Studio · Model Setup

The first time you try H3 Studio, the failure won't be in the nodes. It'll be in the models - this pack deliberately does not download the core MiniMax H3 weights for you, and a missing 32B text encoder is a confusing error to hit at 2am. H3StudioModelSetup is the node that stops you from guessing. It's an action node: no inputs, no outputs, you just run it and read what it tells you.

Its job is to verify the maintained MiniMax H3 model set against what's installed in your ComfyUI folders and report what's missing or mismatched. The description mentions it "uses Universal Asset Downloader when available" - UAD is optional, and the README is explicit that manually installed models work fine. What this node is really doing is answering the question "did I put the right files in the right folders?" before you burn GPU time on a workflow that can't possibly run.

What it checks for

The H3 image stack is more than one file, and this node walks you through all of it:

• the pruned W4A8 FL2VA and REF2VA diffusion transformers in models/diffusion_models/ (from Kijai/MiniMax-H3-experimental)
• the MiniMax H3 32B Qwen text encoder in models/text_encoders/ (Comfy-Org/MiniMax-H3)
• the H3 video VAE in models/vae/

Plus the optional-but-common extras: the Qwen3-VL analyzers/writers, LightX LoRAs, TAEH3 preview model, the experimental T=1 image VAE, and the Face Refine detector/SAM files. If you're following the quick-start, the base three are non-negotiable - the rest are additive and each one degrades gracefully when absent.

Why you should run it once

Because the alternative is reading errors. A missing VAE fails at decode, a missing encoder fails at conditioning, and a missing transformer fails mid-sampling. This node front-loads all of that into one check you can run before you've even wired a prompt, which is exactly the kind of thing that makes alpha software tolerable. It also helps you triage optional features: Face Refine, LightX acceleration, and the Qwen prompt pipeline each have their own fallback behavior, so knowing which optional pieces are actually present tells you which fast paths are live and which are silently falling back.

One thing to keep in mind while you're verifying: this pack is the wrapper, but the models underneath carry their own baggage. MiniMax H3's community license geofences the local weights out of the US, EU, UK and South Korea, and it reaches outputs, not just weights - worth knowing before you go through the trouble of a 40GB+ download and a few hours of setup, wherever you happen to be.

Install

cd ComfyUI/custom_nodes
git clone https://github.com/thaakeno/ComfyUI-MiniMax-H3-Studio.git
cd ComfyUI-MiniMax-H3-Studio
python -m pip install -r requirements.txt

Restart ComfyUI and hard-refresh the frontend, then run this node once to confirm your model folders are right. It's the recommended first step in the maintained workflow - a cheap check before the expensive part.