MiniMax H3 Aicg 模型中转

The problem it solves

MiniMax H3 is a 33B omni-modal video model that turns up as four or five separate files: an FL2VA transformer, usually a REF2VA transformer, a text encoder, and two VAEs - one for picture, one for audio. The AICG3D pack ships its own loader that assembles all of that in one panel, and if you're happy letting it pick from your model folders, use that and skip this page.

AICG3D_H3ModelAdapter exists for the other case. You load the transformer yourself - a quantized GGUF, an fp8 build, a UNETLoader feeding a LoRA and a ModelSampling patch chain - and now the rest of the H3 node family wants a MINIMAX_H3_BUNDLE and has no idea what your graph looks like. The adapter is the handoff: it takes the pieces you loaded and packs them into the bundle format the H3 nodes speak.

That's the whole job. It's a plumbing node, and the reason to care is that it unlocks every quantization and patch chain the ecosystem invents without waiting for the pack author to add a dropdown.

How it works

Mechanically it's boring in the best way. The node builds a MiniMaxH3Bundle dataclass out of the objects you hand it and stamps the internal name fields as "connected" - meaning downstream nodes use the objects directly instead of re-scanning models/ for a filename. No file lookup, no re-load, no surprises about which copy of a VAE you actually grabbed.

Two checks are worth knowing about, because they fire early and they're unambiguous:

• You must connect at least one transformer. The node raises Connect at least one transformer MODEL: FL2VA or REF2VA. if both fl2va_model and ref2va_model are empty.
• It validates the VAE roles. Video VAE and audio VAE are not interchangeable, and the node refuses the pair if you swap them, rather than letting you find out three minutes into a sample.

Because bundling is free, the node reports itself as always-changed to ComfyUI's cache - cheap insurance against a cached bundle pointing at a model you've since replaced.

The inputs that matter

Three required, two optional, and most people only touch the first three:

• text_encoder (CLIP) - the H3 text encoder.
• video_vae (VAE) - the H3 video VAE. If you've been running image models, this is not your ae.safetensors; feeding a normal image VAE in here errors out on load, and that's exactly what you want it to do.
• audio_vae (VAE) - the separate audio VAE (minimax_h3_audio_vae_fp32.safetensors in the pack's own loader dropdown). H3 generates sound natively, which is why there are two.
• fl2va_model / ref2va_model (MODEL, optional) - first/last-frame backbone and multi-reference backbone. Fused or "refdelta"-style models that carry reference ability inside the FL2VA weights can go in fl2va_model alone.

One output: h3_bundle (MINIMAX_H3_BUNDLE), which feeds the pack's MiniMax H3 Aicg context node - the one that actually builds conditioning, latent and the h3_context that renderers consume.

Note what it does not give you: a bare MODEL output. The pack's loader hands out h3_bundle and model because its acceleration widgets patch the model on the way out. The adapter assumes your model is already patched and just wraps it.

Install

Via Manager, search the Registry for AICG3D, 短剧工作台 or MiniMax H3 (package name aicg3d-h3-suite, listed since v1.1.11). Or by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/JGRFW/comfyui-AICG3D

Restart ComfyUI and hit Ctrl+F5 - the pack caches its frontend JS aggressively, and hard-reloading is the difference between "the new panel isn't there" and it working.

requirements.txt is one line: imageio-ffmpeg. Install it into the Python environment ComfyUI is actually running from:

python -m pip install imageio-ffmpeg

You'll also need the H3 weights themselves - transformers into models/diffusion_models, text encoder into models/text_encoders, both VAEs into models/vae. One non-technical snag: the MiniMax H3 Community License excludes the US, EU, UK and South Korea from its Applicable Territory, outputs included. That's a licence question, not an install bug, but it's the reason a download link can look broken from some regions.

Things that go wrong

Both model slots empty. Covered above, and it's the most common first-run error - people wire the VAEs and forget the transformer because the H3 context node has its own model slot.

Swapped VAEs. The video VAE is a 3D latent that carries a temporal axis; the audio VAE is 2D. The adapter's role check catches this at bundle time instead of at the joint decode, which saves you a long wait.

Two H3 packs installed. The adapter's class name is prefixed AICG3D_H3_ on purpose - the pack renames every upstream node into its own namespace at startup so it can't be shadowed by another MiniMax H3 fork. If you also installed a separate prompt-split add-on from the same author, remove it; duplicate node IDs eat each other's frontend logic.