🅛🅣🅧 Set Audio Ref Tokens

LTX-2's party trick is synchronized audio and video in one model, and the thing that genuinely stunned people at launch was voice cloning - one user watched it continue a real video and reproduce the speaker's voice "off of just a few seconds" and posted the very reasonable reaction: "HOW????" This node is how you feed that reference voice in. It's the speaker-identity input for lipdub, dubbing, and any audio generation where you want the output to sound like a specific person.

How it works

The node takes a reference audio latent and attaches it to your conditioning as context. Per its description, the reference tokens are prepended with negative temporal positions - meaning the model reads them as "here's context from before the clip starts," not as frames it needs to generate. So the model gets a clear sense of the target voice and speaking style without trying to recreate that exact audio in the output. That's what keeps the generated voice consistent with the reference while still saying the new lines. This is the mechanism under the Lipdub IC-LoRA workflow, where reference audio tokens are what preserve speaker identity across a dub.

The inputs and outputs

• positive / negative - your existing conditioning, which the node augments with the reference tokens.
• audio_latent - the encoded reference audio (a few seconds of the target speaker is enough).

Outputs are the updated positive and negative conditioning, plus frozen_audio - the reference latent passed along so a two-stage pipeline can keep the audio locked while it upscales the video. Wire the conditioning into your sampler and frozen_audio into the stage that shouldn't touch it.

Installing it

• ComfyUI Manager - search LTXVideo, install, restart.
• Manual - cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo, then restart.

This lives in the audio/IC-LoRA side of LTX-2, so you'll need the LTX-2 model, an audio VAE to encode your reference, and typically the lipdub or dubbing workflow around it.

Common issues

The cloned voice sounds off or generic. Reference quality is everything. A clean, isolated few seconds of the target speaker beats a noisy longer clip. Make sure audio_latent was encoded from good source audio, not a compressed mess.

The audio drifts or gets re-generated during a second stage. That's what frozen_audio is for - route it into the upscale stage so the model freezes the audio and only works on the video, which is exactly how the two-stage lipdub pipeline is designed.

Expectations check. LTX-2's sync is genuinely strong when it behaves, but the community also reports identity drift between shots and background characters randomly starting to lip-sync. Keep clips short and reference audio tight, and don't expect flawless consistency across a long multi-shot piece.