LTX2 Attention Tuner Patch

First thing to know: the node's own description opens with EXPERIMENTAL in all caps, and you should take that at face value. This is a bleeding-edge knob for LTX-2, the audio+video model, and it's the kind of thing Kijai ships within hours of a release so people can start poking at it. It's not a "set and forget" node - it's a research toy that happens to be genuinely useful.

What it does: it swaps in a custom forward pass for LTX-2 that lets you scale the attention per modality. LTX-2 generates video and audio together, and internally those two streams attend to each other. This node exposes four dials for that cross-talk, plus a nice side effect - the description says it also reduces peak VRAM usage, which on a model this heavy is not a small thing.

Why you'd reach for it

LTX-2 is the model everyone's testing because it does synchronized audio and video, and the ComfyUI integration lands fast because Kijai has a track record here - the community's expectation for any LTX release is now "Kijai will have fp8 conversions and helper nodes up before you finish reading the announcement," and that's held across multiple LTX drops. The catch with joint audio/video models is that the two streams can fight: audio bleeds artifacts into motion, or video dominates and the audio goes mushy. This node is the surgical tool for rebalancing that.

The VRAM reduction is the quieter reason to care. LTX-2 is demanding, and anything that lowers the peak lets you fit a longer clip or a higher resolution on the same card.

The inputs and outputs that matter

You wire your model (MODEL) in and get a patched model (MODEL) out - drop the output into your sampler like any other model patch. The dials in between:

• video_scale and audio_scale (both default 1.0) - how strongly each modality attends within itself. 1.0 is the untouched baseline; go up or down from there.
• audio_to_video_scale and video_to_audio_scale (both default 1.0) - the cross-modal terms. This is where the interesting tuning lives: if audio is corrupting your video, pull audio_to_video_scale down; if the audio isn't tracking the picture, nudge video_to_audio_scale up.
• blocks (STRING, empty by default) - which transformer blocks to apply the patch to. Empty means the node's default set. Leave it alone until you know which blocks you want to target.
• triton_kernels (BOOLEAN, on by default) - uses Triton for the custom kernels. Great for speed and VRAM, but it needs a working Triton install.

Change one dial at a time. With four interacting scales, moving several at once means you learn nothing about which one did what.

How to install it

ComfyUI Manager: search KJNodes for ComfyUI, install, restart. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-KJNodes
pip install -r ComfyUI-KJNodes/requirements.txt

Restart, and it appears under KJNodes/ltxv. The pack itself keeps dependencies minimal - but this specific node leans on Triton for the fast path (see below).

Common issues & troubleshooting

Triton errors on Windows. triton_kernels on requires a working Triton, and Triton on Windows portable installs is famously fiddly - it's one of the most common LTX/Wan support threads out there. If you hit an inductor or Triton error at sample time, either fix your Triton install or flip triton_kernels off to fall back to the non-kernel path.

Nothing seems to change. With all four scales at 1.0 the node is a near no-op by design - that's the neutral state. You have to actually move the dials off 1.0 to see an effect, and small moves on a model this sensitive can still be subtle.

It's unstable / results look broken. It says EXPERIMENTAL for a reason. Push a scale too far and the model falls apart. Treat 1.0 as home base, make small excursions, and if a setting nukes your output, that's the tool working as advertised - dial back.