LTX 2.3 AutoDownload (CRT)

LTX 2.3 is not a one-file install. To actually run it you need the 22B transformer, the video VAE, the audio VAE, a text encoder (Gemma 3 12B), a text projection bridge, a preview VAE, and - for the fancy workflows - the union-control LoRA and a spatial upscaler. That's eight separate downloads from four different Hugging Face repos. LTX 2.3 AutoDownload (CRT) is a panel node that turns that chore into a checklist.

This node is not a loader. It has no inputs and no outputs, and it never touches a model itself - the docstring in the source is explicit about it: it's a utility that checks what's missing and downloads it, and you use the normal loader nodes afterward. What it does have is a full JavaScript frontend.

What the panel does

Drop it on the canvas and it renders a list of the eight LTX 2.3 components - Diffusion Model, IC LoRA Union Control, Video VAE, Audio VAE, TAE VAE (preview), CLIP Gemma, CLIP Projection, and Spatial Upscaler. For each row it shows the target filename, the file size, and a status: present, missing, downloading, or error. Click a download button and the model streams down in a background thread while the panel updates progress in place - no freezing your graph, no guessing whether it's working. Behind the scenes it talks to API routes the pack registers under /crt/ltx23/ (check_models, download_model, download_status).

That's genuinely nicer than downloading from the browser. You see exactly which of the eight are missing, each one goes to the correct model subfolder automatically, and partial failures (a dropped connection) clean up the .part temp file so you can retry without a corrupted install.

When to use it vs. the AutoDL loaders

Here's the thing: the pack has two download strategies. The AutoDL loaders (the CRT/AutoDL/LTX2.3 family) download whatever they need when you first run them, invisibly. This panel is for when you want to grab the whole stack up front - the "I'm about to set up a fresh LTX 2.3 machine, let me pull everything now" case. Both approaches write to the same model folders, so they don't fight; the panel just checks the same files the loaders will eventually look for.

Install and gotchas

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

or ComfyUI Manager → search "CRT-Nodes". It's conditionally registered - if the pack fails to import the module, the node simply won't appear (the pack logs a warning and keeps loading everything else). The big honest caveat is disk and download time: the Gemma 3 12B text encoder alone is a ~12GB download, and the whole stack is tens of gigabytes. Check your disk space before you click all eight buttons. Also note this node downloads but never loads - after the panel finishes, you still build your workflow with the AutoDL loaders or the LTX 2.3 US Models Pipe route. If you only care about a couple of files, the individual AutoDL nodes are the leaner path.