MiniMax H3 adaLN Modality

MiniMax H3 generates picture and stereo audio in one pass, and the weights are shared: audio and video both go through the same 50 blocks. That's the point of the model, and it's also why you can't do what LTX 2.3 users do and target "the audio layers" - there aren't any.

There is exactly one pathway that splits cleanly, and this node is a knob for it.

What it does

H3's AdalnProj emits three contiguous row blocks - {video: 0, text: 1, audio: 2} - so the projection is already partitioned by modality. Scaling a slice of a LoRA's lora_B for one of those blocks scales that modality's modulation, with no runtime hook involved. Set video to 1.0 and audio to 0.0 and the LoRA's adaLN contribution stops touching the audio side.

In practice that's the difference between "this style LoRA also changed how my dialogue sounds" and "this style LoRA just restyled the picture". The README's numbers explain why it's worth caring: where adaLN is present in a LoRA, it's 89–99.7% of weight-space perturbation for content LoRAs (median ~96%). Read that as where most of the weight change lives, not as 96% of what you see - but it does mean adaLN is where the LoRA is.

Inputs and outputs

Three floats, no wires in: video, text, audio, each default 1.0 and each spanning 0 to 2 in 0.05 steps. video is tag 0, text is the conditioning modality (tag 1), audio is tag 2.

All three at 1.0 is a declared no-op - the node will tell you so. 0.0 removes that modality's share of the adapter, and going above 1.0 pushes it.

Outputs are adaln_modality, which wires into the stack's adaln_modality input, and info, a short string describing what you set. info is genuinely useful here because it tells you at a glance whether you've actually changed anything.

Two ordering details that matter if you're doing anything clever: scaling runs before adaLN porting, so a .diff_b bias produced by the dense↔curve rebase inherits your modality scaling rather than escaping it. And geometry is read off the loaded AdalnProj; final_layer.adaln_proj is a single-modality projection and is deliberately left alone.

What it won't do

It doesn't isolate a modality, and the author is upfront about that. Attention is joint over the packed sequence, so this changes where a LoRA is applied, not everything it eventually reaches. Don't expect a clean stem separation.

It also only affects LoRAs that carry adaLN pairs. A style LoRA with no adaLN tensors is untouched by any setting here, and the stack's report will mark the control INACTIVE for that row - that's your signal that you're turning a knob attached to nothing.

Install

Ships with the pack. Nothing to install beyond the pack itself, and no Python dependencies.

cd ComfyUI/custom_nodes
git clone https://github.com/cicalooo/ComfyUI-H3-PowerLoraStack

Restart ComfyUI, then find MiniMax H3 adaLN Modality in the MiniMax-H3/lora category. In ComfyUI Manager, search MiniMax H3 Power LoRA Stack.

Common issues

The stack will tell you when this node is doing nothing: check the report output for a per-LoRA line naming the modality values, and the word INACTIVE where a LoRA has no adaLN pairs to scale.

If you're not using the Power LoRA Stack, this node is inert - it produces a data object for one specific input and has nowhere else to go. Which brings up the nice edge case: an empty stack still runs the incoming adaLN pass, so the stack can sit after another loader purely as an AdaLN fix.

And if you're on a curve checkpoint with a dense-trained LoRA, remember the porting node has to do its job before any of this means much - the two features stack in a specific order, and the report shows both.