LTXV Audio VAE Decode

Every LTX-2 workflow ends in the same place: a latent that's secretly a song, a voice, a crowd cheering - and this node is what turns it into something you can actually play. LTXV Audio VAE Decode is the decoder at the end of the audio chain: it takes a sampled audio latent and renders a real waveform, sample rate and all. Without it you have a tensor; with it you have sound you can save to a file or attach to your video.

Two inputs, one output:

• samples - the audio LATENT you got out of the sampler (the same kind of latent LTXV Audio VAE Encode produces, or the one LTXV Empty Latent Audio created as a blank to be filled).
• audio_vae - the audio VAE you loaded. It must be the same VAE that encoded or generated the latent; a mismatched one produces noise, not audio.
• Output: an Audio object - a waveform plus its sample rate - which is exactly what Save Animated WEBM, video-output nodes, and audio preview nodes consume.

Mechanically it's the inverse of encode: the latent gets run through the audio VAE's decoder and normalized (the source applies a loudness normalization step so quiet generations don't come out silent and loud ones don't clip), and the sample rate comes from the VAE itself rather than being a setting you can get wrong.

Where it fits

• Text-to-audio: empty audio latent → sampler → this node → save. LTX-2 does foley, ambient sound, and voice from a prompt, which is the "synchronized audio" selling point in its purest form.
• Audio+video: the model generates video and audio latents together; both get decoded - this node for the sound, VAEDecode for the frames - and a video-output node muxes them into one file.
• Video-to-video sound design: generate audio for an existing clip (the "put sound on my Wan movie" workflow), decode, and composite.

Gotchas

• The VAE must match the checkpoint generation. This is the most common failure: LTX-2.3 introduced a rebuilt VAE with a new latent space, and decoding 2.3 latents with a 2.0/2.1 VAE (or vice versa) gives garbage or an error. Version-match your audio VAE loader to the checkpoint you sampled with. A related LTX-2.3 quirk: the TinyVAE preview decoder stopped working after the VAE rebuild, so you can't preview - this node is your actual hearing.
• Distilled LoRA mumbling. If you're running the distilled LoRA and audio comes out mumbling or mushy, the community fix is the 1.1 revision of that LoRA at strength ~0.4–0.5 - not this node. The decode is faithful; the sampling was the problem.
• Output path: the node hands you audio, but nothing saves it automatically. Wire it into a save node or a video muxer; a bare decode with no consumer just sits there.

It's a small node in a big pipeline, but it's the only place where the hours of setup pay off in something your ears can verify. When the sync is good - and it's genuinely, startlingly good when it works - this node is where you'll be smiling.