MiniMax H3 Easy Loader

The pack's main loader, and why it's not a normal loader

MiniMax H3 isn't one transformer - it ships a text/image/keyframe model (FL2VA) and a full-reference model (Ref2VA), plus a Qwen3-VL text encoder and two VAEs (one for video, one for audio). A plain UNETLoader + CLIPLoader combo doesn't know any of that, which is why this pack has its own. MiniMaxH3EasyLoader selects all five pieces and bundles them into a single MINIMAX_H3_BUNDLE that the MiniMax H3 Easy node consumes. It's the fastest way to a working H3 graph: drop it in, pick files, done.

Inputs

• fl2va_model - the FL2VA transformer dropdown. Pick a .safetensors or .gguf file from ComfyUI/models/diffusion_models/, or leave it as none.
• ref2va_model - same idea for the Ref2VA transformer. You can set one of these to None and the remaining transformer serves every mode. If you only want to test the reference path, that's one download instead of two.
• text_encoder - defaults to qwen3vl_32b_minimax_h3_nvfp4_awq.safetensors (a quantized Qwen3-VL, loaded with ComfyUI's CLIPLoader using the minimax type).
• video_vae - minimax_h3_video_vae_fp16.safetensors.
• audio_vae - minimax_h3_audio_vae_fp32.safetensors.

The output is a single h3_bundle, which you wire straight into the main Easy node. That's the whole job of the node's public face.

How the loading actually works

This is where it gets interesting, because the loader is lazier than it looks. The text encoder and both VAEs load eagerly. The transformers don't - they're loaded on demand by the main node, only when a run needs them. When your mode flips to a different transformer file, the loader releases its cached transformer and calls soft_empty_cache() before loading the other one, so you're not holding both 20 GB models in VRAM just because they're both in the dropdown.

The filename matcher recognizes community naming and quantization variants, including .safetensors and .gguf. If you pick a .gguf transformer or text encoder, the loader routes it through ComfyUI-GGUF's loader nodes (UnetLoaderGGUF / CLIPLoaderGGUF) - and if you haven't installed ComfyUI-GGUF, it raises a clear "install GGUF nodes" error instead of failing cryptically. Regular safetensors use native loaders, no extra pack needed.

Install and models

cd ComfyUI/custom_nodes
git clone https://github.com/nkxx188/ComfyUI-MiniMaxH3-Easy

or search "MiniMax H3 Easy" in ComfyUI Manager, then restart. The Python deps are trivial (requests, psutil); the downloads are the real cost. Standard folder layout:

ComfyUI/models/diffusion_models/   <- FL2VA / Ref2VA transformers
ComfyUI/models/text_encoders/      <- Qwen3-VL
ComfyUI/models/vae/                <- both H3 VAEs

The workflow README points at the community files people actually use: minimax_h3_fl2va_pruned_w4a8_mixed.safetensors from Kijai's experimental repo and the LightX2V MiniMax H3 Turbo LoRA. Full-precision H3 is roughly 42 GB, so the pruned W4A8 and GGUF releases are the realistic path for a single consumer GPU.

Common issues

The usual one is "the model I downloaded isn't in the dropdown." That's almost always a folder problem - transformer files go in diffusion_models/, not checkpoints/, and the dropdown only refreshes on restart. Second: you get a GGUF error because you picked a .gguf file without installing ComfyUI-GGUF. Install it, restart, done. And before you start downloading, remember the H3 Community License excludes the US, EU, UK and South Korea - the weights are geofenced, so verify you're entitled to run them locally before committing to the bandwidth.