Nodes/comfyui-minimax-h3-audio-T8/MiniMax H3 Attention Hooks (T8 Advanced)
ComfyUI Node

MiniMax H3 Attention Hooks (T8 Advanced)

When other packs' attention patch nodes won't talk to H3, this shim translates

By T8mars·Created about a month ago·Updated about 20 hours ago· 1,023
MiniMax H3 Attention Hooks (T8 Advanced)
  • model
  • model
  • report_json

H3 isn't a UNet and it isn't quite a vanilla DiT either - it's a fresh omni architecture with its own attention wiring. That's great for generation and bad for the long tail of community node packs that assume every model exposes the standard attention patch points. When you hook one of those packs onto an H3 model and it silently does nothing, or your whole workflow turns red, the usual culprit is that the standard attn1 patch contract isn't there for them to grab. This node is the compatibility layer that fixes that on older H3 cores.

It's part of the T8 pack's 20-core-compatibility family: nodes that exist to make H3 play nicely with the rest of the ecosystem rather than add new generation tricks. The job here is narrow - expose standard attn1 and attn1-output patch hooks on older H3 cores so external attention-patching nodes (consistency tools, style transfer, any of the packs that patch attention mid-sample) have something to attach to. On newer ComfyUI builds that already ship the official hook contract for H3, the node does nothing - it passes your model straight through unchanged.

How to use it

  • model in → model out, plus a report_json that tells you whether it actually shimmed the hooks or passed through.

Feed the output model into whatever attention-patching node you're trying to satisfy, then continue to the sampler as normal. Wire report_json to a text display if you want to confirm what happened; it's diagnostic, not part of the image path.

The tell for whether you need it at all: if the attention node you want errors with something about missing patch points or hooks on an H3 core, add this between your model loader and that node. If everything already runs, you don't need it - and on current ComfyUI it's effectively a pass-through anyway, so adding it costs you nothing but a wire.

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

The standing gotcha with this pack applies double here: update ComfyUI itself first. It targets recent core APIs (comfy_api.latest, comfy.weight_adapter, comfy.patcher_extension, comfy.ldm.minimax), and an old core makes every T8 node fail import at once - which looks terrifying but is just a version mismatch. requirements.txt is deliberately empty; the base stack comes from ComfyUI, so don't install random optional deps into it.

Where people get burned

The README's warning is worth taking literally: do not stack multiple nodes that each take ownership of the sampler, attention, or MODEL forward path. H3 is sensitive to multiple would-be owners fighting over attention. If you add this shim and another pack's attention wrapper and something else that patches forward, you can end up with slower sampling or weird quality instead of the feature you wanted. Use one attention owner. And remember this node is a compatibility bridge, not a feature - it won't make a broken attention node work if that node simply doesn't support H3's architecture yet.

CategoryT8/MiniMax H3/Compatibility/Advanced

Inputs (1)

NameTypeDefaultDescription
modelMODEL

Outputs (2)

NameTypeDescription
modelMODEL
report_jsonSTRING