LTXV Reference Audio (Ray)

LTXV Reference Audio (Ray) is the node that makes LTX-2/2.3 generations sound like someone specific. LTX is one of the few open-weights models that generates synchronized video and audio, and the community quickly found its audio identity - voices especially - is something you want to steer. This node feeds a reference audio clip into the conditioning and then runs a dedicated identity guidance pass so the generated audio matches the reference's character.

Here's how it works, in two halves. First, the reference: the node takes your reference_audio (an AUDIO object) and runs it through an audio_vae (the LTX audio VAE) to produce audio tokens, resampling the waveform to the VAE's rate if needed. Those tokens get attached to both your positive and negative conditionings as ref_audio - so the model sees the reference during generation. Second, the guidance: the model is patched with a post-CFG function that, for every step in the start_percent–end_percent window, also computes what the model would predict without the reference audio (the conditionings with ref_audio stripped), and pushes the result toward the reference by identity_guidance_scale. That difference is the identity signal - it's how the generation gets steered toward "matches this audio" rather than just "is vaguely audio-consistent."

The inputs

• positive / negative - your standard conditionings. The node returns them back, with ref_audio baked in, so you wire the outputs into the guider, not the originals.
• reference_audio - an AUDIO object (from a Load Audio node).
• audio_vae - the LTX audio VAE (from the pack's LTX Audio VAE loader).
• identity_guidance_scale - how strongly to push toward the reference (default 3). 0 disables the guidance pass entirely, which is a clean way to A/B it.
• start_percent / end_percent - the sampling window where identity guidance applies (defaults 0 to 1 = everywhere). Narrowing the window, say to the first half of sampling, lets the reference shape the early structure without fighting the model later.

It returns ray_actors plus the two conditionings. Because the guidance patch is pushed to every Ray worker, the multi-GPU run behaves identically to single-GPU - this is the raylight-specific replacement for the stock LTX reference-audio setup.

Practical notes

The end_percent window is the lever that matters most. Running identity guidance for the entire sample (default) is strong and can leave audio sounding a bit flat or "too on the nose." If the generated voice feels over-constrained, narrow the window or drop the scale toward 1–2. If it ignores the reference entirely, raise scale and double-check the audio VAE actually produced tokens (a rate mismatch silently resamples, but a missing VAE is an error).

One honest caveat: this steers identity - the timbre and character of the audio. It won't fix LTX's compressed, "tin can" audio quality, which is a model-level limitation the community has documented since LTX-2's launch. Think of it as consistency control, not a quality upgrade.

Install: ComfyUI Manager → "raylight", or clone + pip install -r requirements.txt. You'll need the LTX audio VAE model files, same ones the pack's LTX audio workflows use.