LTXV Audio VAE Encode

The reason people installed LTX-2 is staring at you from this node: give it a voice clip and LTX-2 will produce video synchronized to that voice - lips moving, sound coming out of the right mouth, the whole scarily-good package. LTXV Audio VAE Encode is the front door for that: it takes real audio and compresses it into the latent form LTX-2 conditions on. No audio in, no audio-conditioned generation.

Two inputs, one output:

• audio - an AUDIO tensor, usually from a Load Audio node (or the audio track pulled off a video with the video tools).
• audio_vae - the audio VAE, loaded via the LTXV Audio VAE Loader node. This is the model that defines the latent space, so it must match the one your LTX-2 checkpoint was trained with.
• Output: an Audio Latent (a LATENT with the audio marker) that wires into the sampling side of an LTX-2 workflow.

Mechanically it's simple: the node resamples your audio to the VAE's native sample rate (44100 Hz for the LTX-2 audio VAE) and runs it through the encoder - exactly the same encode that VAE Encode Audio does for images, just the audio flavor. That encoded latent becomes part of the conditioning stack, so when the model generates, it's trying to produce audio+video that matches the reference you fed in. That's how LTX-2 does voice cloning from a few seconds of input, and why people were posting "HOW????" threads about it in the launch week.

Where it fits

• Audio-to-video: encode a song or dialogue track, feed the latent into the sampler alongside an empty audio+video latent pair, and the model generates matching visuals. This is the music-video workflow (10–12 second clips, per community practice).
• Adding sound to existing video: LTX-2's party trick - run video-to-video with your Wan clip and let LTX-2 generate the audio for it, because no other open model can.
• Lipsync and avatars: encode a reference voice and drive a talking head.

Gotchas

• VAE mismatch is the silent killer. If the audio VAE you loaded doesn't match the checkpoint (say, a 0.9.x audio VAE against an LTX-2 model, or one of the pre-2.3 files), the encoded latent is in the wrong space and you'll get garbled sound or a straight error. Load the VAE that came with your checkpoint version.
• Sampling rate mismatch is handled, but quality isn't. The node resamples to 44100 Hz automatically, but a heavily compressed or 8 kHz source stays heavily compressed - garbage in, tinny out. Community consensus on LTX-2 launch audio was "it sounds like a tin can" until you feed it decent source audio.
• No audio track at all? The node errors rather than silently inventing silence (the generic VAE Encode Audio does the same). If your source video has no audio track, that's your problem to solve upstream, not a node setting.

Pair it with LTXV Audio VAE Decode on the other side of the sampler - encode real audio to condition, decode the generated latent back to a playable waveform. Everything between those two nodes is where the magic happens.