🅛🅣🅧 IC-LoRA Loader Model Only

This is the loader that turns on LTX's structural control. IC-LoRAs - In-Context LoRAs - are how LTX takes a control signal like a depth map, pose video, or canny edges and conditions generation on it, and this node is what applies one of those LoRAs to your model. The "Model Only" part means it patches just the MODEL and leaves your CLIP untouched, which is the right shape for video control adapters. If you want depth or pose control on LTX, you start here.

It's part of ComfyUI-LTXVideo, Lightricks' official pack. Lightricks leaned hard into control LoRAs - the 13B shipped official pose/depth/canny adapters, and LTX-2/2.3 added a whole Creative Lab set (union control, motion track, HDR, colorization, and more). This loader is the common gateway to all of them.

How it works

It loads a LoRA and applies it to the model at a chosen strength, like any LoRA loader - but with one LTX-specific extra. IC-LoRAs are often trained to operate on a downsampled latent grid to save memory and speed up inference, and the loader reads that latent_downscale_factor straight out of the safetensors metadata and exposes it as an output. That number then feeds the guide node that injects your control video, so the two agree on scale automatically instead of you hand-matching it.

The inputs and outputs that matter

• model - the LTX model to patch.
• lora_name - pick your IC-LoRA from models/loras. Empty dropdown means the file isn't there.
• strength_model - how hard the LoRA applies (default 1). Lower it for a lighter touch; the Lightricks union LoRAs in particular are tuned around specific strengths.

Outputs are the patched model and a latent_downscale_factor (a FLOAT). Wire the model onward to your sampler/guider, and route the downscale factor into the matching input on an IC-LoRA guide node (like Add Video IC-LoRA Guide), so the control frames get encoded at the resolution the LoRA expects.

Installing it

ComfyUI Manager → Install Custom Nodes → search LTXVideo → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/Lightricks/ComfyUI-LTXVideo and restart. It's under Lightricks/IC-LoRA. The IC-LoRA files themselves are separate downloads - the README lists them (union control, motion track, HDR, pose, detailer, and the rest) on Lightricks' HuggingFace; drop them in models/loras.

Where people get burned

• Ignoring the downscale factor. The whole reason this node outputs latent_downscale_factor is so your guide node encodes the control video at the right scale. Leave it unconnected and hard-code a mismatched value, and the control signal lands at the wrong resolution - soft or broken guidance.
• Wrong version LoRA. LTX moved fast: 0.9.7 control LoRAs, LTX-2 (19b) adapters, and LTX-2.3 (22b) adapters aren't interchangeable. A "LTX-2" labeled file may not be a 2.3 file - match the LoRA generation to your checkpoint, because docs and filenames both blur "LTX-2" and "2.3."
• Strength too high. Overdriving strength_model can make the control dominate and the prompt vanish. Start at the value the LoRA's page recommends (the ref0.5 union LoRAs, for instance, are built around a specific setup) and adjust.
• Expecting it to condition by itself. This loader only patches the model. You still need a guide node to feed in the actual control frames - the loader and the guide are a pair.