LTX2.5 VIDEO VAE (CRT AutoDL)

LTX2.5 VIDEO VAE (CRT AutoDL) is the frame-decoding half of LTX2.5's two-VAE design. No inputs, one VAE output, and on first run it downloads ltx-2.5-video-vae-bf16.safetensors into ComfyUI/models/vae/ from the CRT_Nodes_25Gate repo. That's the whole node - and for the AutoDL family, that's the point.

LTX-2.5 carries on the LTX-2 architecture's most distinctive trait: separate video and audio VAEs, because the model generates picture and synchronized sound in one pass. You can't decode an LTX2.5 latent with a generic VAE - the latent space is the "new latent space" that the 2.5 generation was built around, and the VAE is the matching key. This node hands you the right one, in bf16, no file-hunting.

Where it fits

In a complete LTX2.5 graph, the video VAE decodes the frame latent into images/video while LTX2.5 AUDIO VAE decodes the audio latent into a waveform. Both come from the same CRT stack, and both are needed if you're generating the synchronized-audio side. For a silent-video run you can skip the audio VAE, but the video one is never optional.

Output is a VAE that plugs into the decode step of your LTX2.5 sampler. Because it's bf16 it's a lighter download than fp32 audio VAE, and because it's a fixed target it's the same file every time - no version roulette.

Installing it

Standard CRT install:

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. First run downloads the VAE (a couple of GB), then it behaves like any other VAE loader.

Common issues

• Wrong VAE gives garbage - this decodes LTX2.5 video latents only. Pair it with the LTX2.5 Model and the matching AUDIO VAE; don't reuse an LTX-2.3 VAE or a generic one.
• Audio missing - you've decoded video but not audio. Add LTX2.5 AUDIO VAE to the audio-latent side.
• Update breakage - red socket or NaN → right-click → Fix node (recreate).