Empty Latent Audio

Video generation stopped being silent a while back, and silent-empty-latents became a whole sub-genre. This node allocates a latent for audio - the compressed sound track that gets sampled alongside the video frames. If your workflow has a model with a joint audio-video pipeline, this is the box you hand the sampler so it has somewhere to put the sound.

What it creates

The node takes seconds (default 47.6, up to 1000) and batch_size, and produces [batch_size, 64, length] where length is seconds × 44100 / 2048 (rounded to even). The important metadata it attaches is type: "audio" and downscale_ratio_temporal: 2048 - that 2048 is the audio VAE's temporal compression: 2048 raw samples of 44.1kHz audio collapse to one latent step. So a 47.6-second clip becomes a 1024-step latent at 64 channels - the default is chosen precisely because it lands on a clean power of two. That's a heavily compressed, heavily channeled representation - audio latents are wide, unlike image latents.

That 2048 downscale is exactly the family that models like the LTX/Wan-style audio pipelines use, and VAEDecodeAudio is the node on the far side that turns the sampled latent back into sound. The relationship should feel familiar: empty latent → sampler → decode, just with audio shapes instead of video shapes.

Where it sits

You feed it into a sampler that supports the audio latent format alongside your video latent. Because the model generates both at once, your workflow typically holds two latents - one video, one audio - and this is the audio side's empty canvas. Batch size matches your video batch; if the numbers disagree, downstream nodes will complain loudly.

What matters in practice

seconds is the knob you actually think about, and it's a real constraint: the sampler has to denoise every one of those latent steps, and a 60-second clip is a lot of steps. Don't assume audio generation is cheaper than video just because it's invisible - it runs alongside it and shares the same hardware. The 47.6-second default is oddly specific because it's exactly 1024 latent steps at the default sample rate - a clean power of two that happens to be a nice balance between "short demo" and "actually usable clip."

Gotchas

Audio latent shapes are rigid. The length is derived from seconds, so if you want a specific latent length, back-solve from the 2048 ratio rather than guessing seconds and squinting at the result. And the usual rule: this latent is only valid for the audio VAE it's shaped for - a 64-channel, 2048-downscale latent is not interchangeable with the AceStep family's audio latents, which use different ratios and shapes entirely. Match the empty latent to the model family, and you're fine.