LTX2 Audio Latent Normalizing Sampling(LTX2AudioLatentNormalizingSampling)

Description

The LTX2 Audio Latent Normalizing Sampling node improves the quality of audio generated by LTX2 models by applying per-step normalization to the audio latents during the sampling process. This can help reduce artifacts and improve coherence, particularly for longer audio generations.

Inputs

• model (MODEL, required)
  The LTX2 model to which the audio normalization override will be added.

• audio_normalization_factors (STRING, required, default: "1,1,0.25,1,1,0.25,1,1")
  A comma-separated list of normalization factors, one per sampling step. Each factor scales the audio latent at that step.
  For example, the default "1,1,0.25,1,1,0.25,1,1" applies a factor of 0.25 at steps 3 and 6 (counting from 1), attenuating the latent at those steps. Factors can be any positive float; values less than 1 reduce the latent magnitude, values greater than 1 amplify it.
  The list length should match the total number of sampling steps used in the generation pipeline. If it is longer, extra values are ignored; if shorter, the last value is repeated for remaining steps.

Outputs

• MODEL
  Returns a modified model with the normalization behavior attached. Pass this model to a standard sampling workflow; the normalization factors will be applied automatically during each sampling step.

Usage Notes

• This node is intended for use with LTX2 audio models in the KJNodes/ltxv category.
• The normalization factors are applied in addition to any default latent scaling already performed by the sampler. They act as a per‑step gain adjustment on the audio latent.
• Experimenting with different factor sequences can help reduce high‑frequency noise, stabilize volume, or emphasize certain temporal regions.
• If you are unsure which factors to use, the default sequence (1,1,0.25,1,1,0.25,1,1) is a good starting point for a typical 8‑step sampling schedule. For longer sequences, adjust the pattern accordingly.
• The node does not change the total number of steps; it only overrides the audio latent behavior during the sampler’s iterations.