H3 Studio · Model Loader

MiniMax H3 isn't a model you load with one dropdown. A working H3 setup means a pruned diffusion transformer, a 32B Qwen text/vision encoder, and a video VAE - and if you're using H3 Studio's reference workflows you can be juggling two transformers (FL2VA and REF2VA) that get swapped depending on which route you're on. That's the problem this node exists to hide.

H3StudioLoader is the pack's single loading point. You point it at your model filenames once, and it hands you a typed h3_bundle plus the clip and video_vae outputs the rest of the graph expects. It's the node at the top of the maintained H3_Studio_Unified_Image.json workflow, and honestly, if you're building your own H3 Studio graph, you start here.

What it loads

The inputs are all dropdowns populated from your ComfyUI model folders, so no typing filenames by hand:

• fl2va_model / ref2va_model - your H3 diffusion transformers from models/diffusion_models/ (the Kijai pruned W4A8 .safetensors files from MiniMax-H3-experimental are the defaults). FL2VA handles text-to-image and image-to-image; REF2VA handles reference editing.
• text_encoder - the MiniMax H3 Qwen3-VL 32B conditioning encoder, defaulting to the NVFP4 AWQ build, which is the official ComfyUI template choice. The tooltip warns the INT8 ConvRot variant "can stream very slow", and that's not marketing - that one is genuinely painful.
• video_vae - the native MiniMax H3 video VAE (minimax_h3_video_vae_fp16.safetensors).
• image_vae - optional Mamad8 T=1 image decoder. Experimental and image-only; it never replaces the normal H3 video VAE, it's a lighter one-frame alternative you opt into.
• image_analyzer / prompt_writer - optional Qwen3-VL-family models for the automatic reference analysis and prompt writing. "Same as image analyzer" for the writer reuses one loaded checkpoint, which is the sensible default.

The clever bit is in the mechanism: the transformers aren't loaded at node-execution time. The loader stores their names in the bundle and loads them lazily when the chosen route actually needs one, then releases the previous transformer and asks for a soft cache cleanup when you switch. You pay for one model residency at a time, not both.

The rest

The node outputs h3_bundle, clip, video_vae, and a model_info string. Wire the bundle and clip into H3 Studio · Condition & Route, and the VAE into decode.

One honest gotcha: nothing downloads automatically. The README is explicit that H3 Studio does not fetch the core H3 models for you - that's what the H3 Studio · Model Setup node is for (or the manual links in the README). You also need to have actually installed H3 itself, and worth remembering: the MiniMax H3 Community License geofences the local weights out of the US, EU, UK and South Korea, so check your region before you invest the disk space.

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. The core pack has no dependencies beyond ComfyUI itself; only the optional local VLM analyzers need pip install -e ".[vlm]". Keep in mind this is alpha software, and ComfyUI Nodes 2.0 isn't supported yet - the pack targets classic Nodes 1.0.