MiniMax H3 Fun ControlNet Loader (T8 Advanced)

ControlNet is the one conditioning trick every new base model inherits, and MiniMax H3 is no exception - it just needed its own implementation. The H3 "Fun ControlNet" is a separate set of experimental weights that let you steer what the model generates using a preprocessed video of depth, pose or edges. This loader is the half of that feature that gets the weights off disk. It's deliberately boring, and that's the point: the interesting decisions happen in its partner node, the Fun ControlNet Apply.

It reads from the standard models/controlnet folder, so it feels like every other ControlNet loader you've used. The difference is inside: H3's ControlNet is not a copied UNet encoder the way the SD-era recipe was, and it doesn't load through ComfyUI's generic ControlNet path on older cores. The node uses the official ComfyUI implementation when your build has it, and falls back to a clone-scoped loader when it doesn't - which is the reason you'd reach for it instead of hoping a generic loader works.

How to use it

One input, control_net_name - a dropdown of whatever H3 Fun ControlNet files are in models/controlnet. Outputs:

• control_net (an H3_T8_FUN_CONTROL object) - wire this into the Fun ControlNet Apply node's control_net input
• report_json - which path it used (official or fallback); diagnostic

There's no strength here, no start/end percent - those live on the Apply node. The loader's whole job is "pick the file, load it in a way that works."

The actual model files come from Kijai's MiniMax-H3-experimental HuggingFace repo (controlnet folder) - drop them into ComfyUI/models/controlnet. If the dropdown is empty, that's the problem: you haven't downloaded the weights yet, not a broken node.

Installing it

Part of the T8mars/comfyui-minimax-h3-audio-T8 pack - search "MiniMax H3 Audio T8" in ComfyUI Manager, install, restart, or:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/comfyui-minimax-h3-audio-T8.git minimax-h3-audio-T8

Update ComfyUI itself first - the pack targets recent core APIs (comfy_api.latest, comfy.weight_adapter, comfy.patcher_extension, comfy.ldm.minimax) and fails wholesale on old ones. requirements.txt is intentionally empty; the base stack comes from ComfyUI's own environment.

Where people get burned

Two traps, both from the README. First, don't mix base variants - the Fun ControlNet was validated against particular H3 bases, so pair it with the workflow's declared base (FL2VA/Ref2VA/full) and don't swap. Second, remember the no-stacking rule: this is part of the control/conditioning layer, so don't also pile on other nodes that want to own the attention or MODEL forward path. And keep expectations calibrated - Fun Control is labeled experimental even by the pack's own standards, so treat it as "steer the composition" rather than pixel-perfect alignment. The start/end_percent defaults on the Apply node (guide composition, then release) are exactly the discipline the KB's ControlNet essay recommends: hold structure early, let detail breathe late.