LTX2.3 VIDEO VAE (CRT AutoDL)

Every LTX 2.3 workflow needs a video VAE, and it's the same file every time: the rebuilt LTX23_video_vae_bf16.safetensors from Kijai's ComfyUI repack. This node's entire job is to make sure that file exists and hand it to you as a ready VAE. No inputs, no knobs, one output.

That sounds trivial, but it's the difference between a workflow that works on a fresh machine and one that fails with "VAE file not found" while you hunt HuggingFace for the right download. This is the plumbing of the pack's AutoDL philosophy: a small fleet of zero-config loaders that each own one model file, download it into the correct ComfyUI folder on first use, and then behave like a normal loader forever after.

How it works

Run it once and it checks whether models/vae/ already holds LTX23_video_vae_bf16.safetensors. If not, it pulls it from Kijai's LTX2.3_comfy repo with a progress readout and loads it. The output VAE plugs into the LTX 2.3 Models Pipe alongside the model, the audio VAE, and the text encoders - or directly into a VAE Decode if you're assembling the graph by hand.

LTX 2.3's "rebuilt VAE, new latent space" is a real thing, by the way - that's the change that sharpened fine detail and killed the soft hair/edge look of 2.x. It also means the 2.x-era TinyVAE preview trick stopped working because the preview decoder was built for the old latent space. Don't reach for an old LTX VAE as a substitute; the 2.3 VAE and 2.3 model expect each other.

Installing it

CRT-Nodes install as usual - ComfyUI Manager → CRT-Nodes, or clone + pip install -r requirements.txt, restart. Then drop this node in your graph; the download happens on first execution. It's a few hundred MB, so the first run takes a beat.

Where people get burned

The main trap is expecting it to work with a 2.2 or 0.9.x model. LTX 2.3 is its own latent space; using the 2.3 video VAE against an older checkpoint produces garbage, and vice versa. If your colors come out wrong or your output looks structurally broken, check that every LTX piece in the graph is 2.3. And remember there are two VAEs in a full LTX 2.3 setup - this video one for the frames and a separate audio VAE for the soundtrack. This node is only the video half.