LTX2.5 AUDIO VAE (CRT AutoDL)
The audio half of LTX2.5's split VAE, downloaded for you
- VAE
Since LTX-2, Lightricks' video models generate picture and synchronized audio in one pass - and that means they need a VAE that knows how to turn an audio latent back into a waveform. LTX2.5 AUDIO VAE (CRT AutoDL) is that node: no inputs, one VAE output, and it auto-downloads ltx-2.5-audio-vae-bf16.safetensors into ComfyUI/models/vae/ on first use.
It's one of the "two VAEs" that the LTX-2.5 stack splits the decode across: the VIDEO VAE handles frames, this one handles sound. If the model you're running generates speech, music, or foley alongside the picture - LTX2.5's whole party trick - you need both in the graph. Silent clip? The video VAE covers you; the audio VAE sits unused.
Where it fits
In a CRT LTX2.5 pipeline: model from LTX2.5 Model, text from LTX2.5 CLIP, frames decoded with the video VAE, and the audio latent decoded through this audio VAE. Out the other side you get an AUDIO tensor you can preview, or save with Save Audio With Path (CRT) to keep the soundtrack as its own file. The bf16 dtype keeps the download and memory footprint modest compared to fp32 audio VAEs elsewhere in the pack.
Because it's an AutoDL node, there's no configuration to get wrong: add it, let the first run fetch the weights, and it behaves like any other VAE loader. The only real decision is remembering to actually use it when your workflow has audio.
Installing it
Same as every CRT node - install the pack once:
cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt
or ComfyUI Manager → search CRT-Nodes, restart. The first-run download is a couple of GB; the pack's heavy requirements install is the slower part.
Common issues
- No sound on decode - the audio latent needs this VAE, not the video one. If you've got an audio track in the latent but only wired the video VAE, you'll decode a silent video.
- Wrong latent, wrong VAE - this is LTX2.5's audio VAE. Using it on another model's latents is a guaranteed mess; keep it inside the LTX2.5 stack.
- Update breakage - red socket or
NaN→ right-click → Fix node (recreate).
Inputs (0)
No inputs
Outputs (1)
| Name | Type | Description |
|---|---|---|
| VAE | VAE | — |