🅛🅣🅧 Low VRAM Audio VAE Loader

LTX-2 is a joint audio-video model, and the audio has its own VAE - a separate decoder that turns the audio latent into an actual waveform. This node loads it, and the "Low VRAM" part isn't marketing: it's built to load in sequence with your other big models so you don't try to hold everything on the GPU at once and fall over. On a stack that already includes a 22GB text encoder and a 20B-plus video model, every model you can load in order instead of all together is the difference between fitting and OOM.

How it works

It loads an LTXV Audio VAE checkpoint and returns it as an audio_vae (a VAE type). The trick is the optional dependencies input, described plainly by the node: connect it to a previous loader's output to force sequential loading and reduce peak VRAM. So you chain your loaders - video model, then text encoder, then audio VAE - through this input, and each waits its turn instead of all grabbing memory simultaneously. Downstream, the audio VAE feeds the audio decode node that pulls the generated sound out of the joint latent, which you then save with a standard audio node.

The inputs and output

• ckpt_name - which audio VAE checkpoint to load, from what's in your models folder.
• dependencies (optional) - the sequential-loading hookup. Wire a prior loader here when VRAM is tight; leave it empty if you've got headroom.

Output is audio_vae, a VAE you route into the LTX audio decode.

Installing it

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

You need the LTX-2 audio VAE weights on disk for the dropdown to show anything - the audio side is a distinct download from the main video checkpoint.

Common issues

ckpt_name dropdown is empty. The audio VAE weights aren't installed. Download the LTX-2 audio VAE and drop it in the appropriate models folder, then refresh.

OOM the moment audio loads. You skipped the dependencies chain, so the audio VAE is trying to load alongside everything else. Wire it through a previous loader's output to serialize the loads - that's the entire reason this "low VRAM" variant exists. The --reserve-vram flag helps too.

No sound in the output. This node only loads the decoder; it doesn't decode. Make sure the audio_vae actually feeds an audio decode node and that your workflow is generating audio in the first place - a video-only LTX run has nothing for it to do.