LTX2.3 IC Cnet LoRA (CRT AutoDL)

If you've been around video models a while, you'll instinctively reach for "ControlNet" when you want to steer a generation with depth, pose or edges. LTX doesn't speak that language - Lightricks' control mechanism for 2.x is the IC-LoRA (InContext LoRA), an adapter trained to condition generation on a control input. This node loads LTX 2.3's union control IC-LoRA for you, downloads it on first use, and applies it to your model.

The "AutoDL" in the name is the whole value proposition. Rather than hunting HuggingFace for the right .safetensors and remembering which folder it belongs in, you drop this node in, it fetches the file into the right ComfyUI directory with a progress readout, and applies it. Same trick as the pack's other LTX 2.3 AutoDL loaders - they're the plumbing so the Unified Sampler can just work.

What it loads and does

The file is ltx-2.3-22b-ic-lora-union-control-ref0.5.safetensors from Lightricks' own repo - the official union control adapter for the 22B model, trained so it can consume control references (depth and friends) and steer generation accordingly. You wire a base MODEL in, set strength_model (1.0 default, and it does go negative-to-100 if you want to experiment), and get back:

• MODEL - your model with the LoRA applied.
• latent_downscale_factor (FLOAT) - this is the interesting second output. It reads the LoRA's reference_downscale_factor metadata and passes it out, which tells downstream nodes how much to downscale a reference image before feeding it as control. If the metadata is missing it falls back to 1.0 with a warning. Treat that float as part of the control setup, not an afterthought - feed it where your V2V/reference node expects it.

If you set strength to exactly 0 it returns the model unchanged rather than burning time applying nothing, which is a nice touch.

Installing it

Standard CRT-Nodes install - ComfyUI Manager → CRT-Nodes, or clone plus pip install -r requirements.txt, restart. This node is conditional on the AutoDL machinery loading; it uses the pack's built-in HTTP downloader, so the first run will pull roughly a couple of GB from HuggingFace and drop it in models/loras/. Have a little patience on first use.

Where people get burned

The classic confusion is treating this like a style LoRA - it's not for look, it's for control. If you apply it without any control reference plumbing, you're just adding a LoRA that expects inputs you're not giving it. Pair it with the LTX 2.3 US Config / Unified Sampler setup that knows how to build the depth or reference guide. And remember it's union control - one adapter handling multiple control types - so check which control mode your workflow actually feeds it; the sampler's v2v_mode dropdown (Depth Control, Outpaint, Upscale) is where that decision lives.