LTXV Empty Latent Audio

LTX-2 is the one open video model that spits out synchronized audio and video in a single pass, and this node is how you give it a blank audio track to fill. LTXV Empty Latent Audio is the audio-world equivalent of Empty Latent Image: a tensor of zeros that sampling turns into sound. If you've ever set up a text-to-audio or audio+video workflow and wondered where the audio side "starts," it starts here.

Here's the picture. The LTX-2 model is multimodal: its diffusion transformer denoises both a video latent and an audio latent at the same time, guided by one text encoder (the infamous Gemma 3 12B). The audio latent isn't a waveform you can play - it's a compressed representation in a frequency domain, shaped by the audio VAE that ships with LTX-2. This node builds the empty latent that the sampler then fills, using the audio VAE purely for its configuration.

Four inputs, and the first three are the ones you'll actually touch:

• frames_number - default 97, the number of video frames the audio is being generated for. LTX conventions run around 97 or 121 (that's roughly 4–5 seconds at 24–25 fps).
• frame_rate - default 25, frames per second. Set it to match your video side (24 is the common LTX-2 default, 48 for image-to-video).
• batch_size - default 1, how many audio latents you want in one shot.
• audio_vae - the audio VAE you loaded with the LTXV Audio VAE Loader node. This is required - the node reads the VAE's latent channels and frequency-bin count to size the tensor, so it can't work without it.

The output is a single LATENT with a "type": "audio" marker inside, which tells the sampling nodes this is audio rather than video. That marker matters: LTX-2's sampling nodes treat audio and video latents differently, and they need to know which is which.

Where it fits

Two common graphs. Pure text-to-audio: this node → KSampler with an LTX audio conditioning → LTXV Audio VAE Decode → your ears. Audio+video: the audio latent and an empty video latent (from EmptyLTXVLatentVideo) get fed to the sampler together, and the model generates both - which is the lipsync and music-video superpower that pulled people to LTX-2 in the first place. Community practice for music videos is 10–12 second clips with a 2-second filler head so motion and sound get a chance to establish.

The gotchas

• frames_number and frame_rate are a pair. The node derives how many audio latents it needs from both - there's no standalone "audio duration" field. If your video is 121 frames at 24 fps, set the same numbers here or the two sides won't line up and you'll get drift or a cut-off track.
• It's the right default, not a magic number. 97 frames at 25 fps is the LTX-2 starting point; the model card and community threads are full of people who never touched it and wondered why audio and video ended at different lengths. Make them match.
• LTX-2.3 changed the audio pipeline subtly (new vocoder, cleaner audio), but the node itself is unchanged - it just feeds a latent. If your audio comes out mumbly, that's the distilled LoRA version or sampling, not this node.

The audio VAE lives in your checkpoints folder and is loaded via the LTXV Audio VAE Loader node - it's a separate file from the main model, so a "missing VAE" error here usually means that file never got downloaded, not a wiring problem.