MiniMax H3 Easy Model Adapter

The loader you use when you don't want their loader

The bundled MiniMax H3 Easy Loader is convenient, but it decides how models load: text encoder and both VAEs up front, transformers on demand, GGUF routing through ComfyUI-GGUF. That's the right default for most people. It's not the right tool for everyone. If you're already loading your H3 transformer with a native UNETLoader, a GGUF loader node, a bnb-quantized loader, or anything else in the ecosystem - or if you want to keep models loaded in a specific way - the bundled loader is in your way.

MiniMaxH3EasyModelAdapter is the compat layer that fixes that. It takes ordinary ComfyUI MODEL, CLIP and VAE outputs and reassembles them into the same MINIMAX_H3_BUNDLE the rest of the pack expects. You keep your favorite loaders; the Easy nodes keep working.

Inputs

• text_encoder (CLIP, required) - your Qwen3-VL H3 text encoder, from any CLIP loader that can emit it.
• video_vae (VAE, required) and audio_vae (VAE, required) - both H3 VAEs, from whatever VAE loader you like.
• fl2va_model (MODEL, optional) - the FL2VA transformer.
• ref2va_model (MODEL, optional) - the Ref2VA transformer.

You must connect at least one transformer; connecting both is fine. If you leave one out, the main node falls back to whichever transformer you did connect, same as the bundled loader's single-model behavior.

Output

• h3_bundle (MINIMAX_H3_BUNDLE) - wire into MiniMax H3 Easy's h3_bundle input. From there the graph is identical to the bundled-loader path.

How it differs under the hood

Two things worth knowing. First, the adapter marks itself always-dirty (IS_CHANGED returns NaN), so it re-runs every execution rather than letting ComfyUI's cache skip it. That's deliberate: the node's content is whatever upstream loaders fed it, and the adapter can't know if you swapped models without re-running. The cost is that it defeats caching for everything behind it, so if your workflow feels slower than it should, this is the usual suspect.

Second, memory. The bundled loader swaps transformers on demand and asks ComfyUI to flush the soft cache between mode changes. The adapter can't do that - whatever your upstream loaders loaded stays resident, and if you connect both transformers they both sit in VRAM per ComfyUI's model-management behavior. If you're squeezing a consumer card, connect only the one transformer you're actually using, or keep using the bundled loader.

Install and use

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

or search "MiniMax H3 Easy" in ComfyUI Manager, then restart. No extra dependencies. One practical note: this is also the node to reach for when you want a GGUF transformer but don't want the bundled loader's GGUF routing - load it with whatever GGUF loader you already use, and the adapter will happily bundle the result. If you connect neither transformer you get a clear error, so the common "nothing loaded" mystery is hard to hit.

Should you use it?

If you're new to H3, start with the bundled loader - fewer moving parts, and its on-demand swapping is genuinely friendlier to VRAM. Reach for the adapter when you have a reason to own the loading yourself: a GGUF or quantized pipeline you already trust, a loader the bundled node can't express, or a workflow where you're sharing already-loaded models across graphs.