LTX2 Audio Latent Normalizing Sampling

LTX-2 is one of the video models that generates sound along with the picture - and its audio, out of the box, can come out rough. LTX2AudioLatentNormalizingSampling is a targeted fix for that. It normalizes the audio latents at specific steps during sampling, which smooths out the generated audio and improves its quality. If you're doing LTX-2 video-with-audio and the soundtrack sounds distorted, harsh, or just off, this is the node aimed squarely at that problem. Kijai does day-one fixes for the LTX releases as a matter of routine, and this is that instinct applied to the audio side.

How it works

LTX-2 produces audio and video together in latent space, denoising both over the sampling process. The audio latents can drift in magnitude in ways that hurt the final sound. This node patches the model so that, at chosen sampling steps, the audio latents get renormalized - scaled back toward a sane range - before sampling continues. The result is cleaner audio without touching the video side.

The inputs

There are only two, and the second is the interesting one:

model - your LTX-2 model. The node returns a modified MODEL that you sample with.
audio_normalization_factors - a comma-separated string of numbers, defaulting to 1,1,0.25,1,1,0.25,1,1. Each value is the normalization applied at a step in the schedule. Notice the pattern: mostly 1 (leave it alone) with a couple of 0.25 values (pull it down hard) at specific points. That's the tuned recipe. Unless you're chasing a specific artifact, the default is your starting point - it encodes where in the schedule the audio wants reining in.

The single output is the patched MODEL.

When to reach for it

Only when you're generating audio with LTX-2 and the sound quality is the problem. It does nothing for the video itself and nothing for LTX workflows that don't generate audio. It's a specialist tool for a specialist annoyance - but if that annoyance is yours, it's exactly the right one.

Installing KJNodes

ComfyUI Manager: Custom Nodes Manager, search "KJNodes for ComfyUI", install, restart. Manual: cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-KJNodes, then pip install -r ComfyUI-KJNodes/requirements.txt (portable: python_embeded\python.exe -m pip install -r ComfyUI\custom_nodes\ComfyUI-KJNodes\requirements.txt), restart. The node adds nothing to download; you supply the LTX-2 model stack separately.

Troubleshooting

If it seems to do nothing, make sure the node's output model is what you're actually sampling with - patch nodes like this are useless if you leave the sampler wired to the original model. If the audio gets worse after you edited the factors, revert to the default string; those specific 0.25 values at those specific positions are tuned, and randomly changing them tends to make things worse, not better. And remember the scope: this is an audio-quality tool. If your video is the thing that looks bad, this node is the wrong lever entirely.

Name	Type	Default	Description
model	MODEL		The model to add preview override to.
audio_normalization_factors	STRING	1,1,0.25,1,1,0.25,1,1	Comma-separated list of audio normalization factors to apply at each sampling step. For example, '1,1,0.25,1,1,0.25,1,1' will apply a factor of 0.25 at the 3rd and 6th steps.