LTX2.3 AUDIO VAE (CRT AutoDL)

No widgets, one job: fetch the audio VAE that LTX-2.3 needs and hand you back a VAE. This is the odd one out among LTX's usual loader set, because LTX-2.3 isn't a plain video model - since LTX-2, Lightricks' line generates synchronized audio and video out of the same pass, and that means it needs an audio VAE alongside the usual video one to decode the sound it generates.

Why LTX needs a second VAE at all

LTX-2.3 is Lightricks' 22B update to the LTX line, current as of this writing, with a rebuilt VAE and better prompt adherence and audio quality than LTX-2 before it. The headline capability across the whole LTX-2/2.3 generation is native synchronized audio - lipsync, music video, sound effects generated in the same pass as the picture, not bolted on afterward. That's genuinely rare among local video models, and it's also why the pipeline has more moving parts than a silent video model: video needs decoding through a video VAE, and the audio track that comes out of the same generation needs its own decode path through an audio-specific VAE. This node exists to get that second piece onto your disk without you having to track down which exact file Lightricks ships for it.

Two practical things worth knowing about LTX-2.3 specifically, since they'll bite you elsewhere in the same pipeline even though they don't affect this node directly: width and height must be divisible by 32 and frame count must be divisible by 8 plus 1 - off-grid values fail outright rather than silently rounding - and if you're running the distilled LoRA, use version 1.1, not 1.0, since 1.0 is reported to produce mumbling audio out of the first sampler stage.

The inputs and outputs that matter

• No required or optional inputs - nothing to configure.
• Output: a single VAE.

That's the entire node. Like the rest of the AutoDL family, there's no dropdown for which specific checkpoint or precision to fetch - it's a fixed, one-purpose loader.

Installing it

Ships with the rest of CRT-Nodes; no separate install needed for this node specifically. Per the pack's README, the broader LTX 2.3 AutoDownload machinery in CRT-Nodes "uses built-in HTTP download utilities and adds API routes for model status/download" - this VAE loader is part of that same system, just scoped to one file.

ComfyUI Manager - search CRT-Nodes, install, restart.

Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r requirements.txt

Restart ComfyUI. The audio VAE weights fetch on first run of the node rather than at install time - a smaller download than the LTX diffusion model itself, but still worth expecting a pause for on first use.

Common issues

• Generated video has no audio, or audio decode errors out - check this node is actually wired into your pipeline's audio-decode step, separate from whatever's handling the video VAE decode. It's easy to build an LTX-2.3 workflow around the video path and forget the audio side needs its own VAE entirely.
• First run pauses - that's the download; check your console for HTTP activity.
• Confusing this with LTX's video VAE - this node only produces the audio-side VAE. If your workflow is failing to decode the video frames themselves, the problem is elsewhere in your graph, not here.
• Resolution or frame-count errors elsewhere in the pipeline - not this node's fault, but a very common LTX-2.3 trap worth knowing about while you're setting up the rest of the workflow: width/height need to be divisible by 32 and frame count divisible by 8 plus 1, or the run fails outright rather than adjusting automatically.