LTX2.3 IC Outpaint LoRA (CRT AutoDL)

The LTX 2.3 workflow scene is built around a family of IC-LoRAs - small adapters that teach the base 22B model new behaviors like reference-image control, upscaling, and outpainting. The outpainting one lives on Hugging Face under oumoumad/LTX-2.3-22b-IC-LoRA-Outpaint, and if you've ever manually hunted for the right safetensors URL for a LoRA, you know this node is saving you a chore. LTX2.3 IC Outpaint LoRA (CRT AutoDL) downloads it into models/loras on first use and applies it to your model with a strength you control.

What it does

It's a plain _FixedLoRALoader under the hood. Give it a model and a strength_model (a float, default 1.0, range −100 to 100), and on first run it ensures ltx-2.3-22b-ic-lora-outpaint.safetensors is sitting in your loras folder, then loads it and applies it via ComfyUI's load_lora_for_models. Out comes a MODEL with the LoRA baked in.

strength_model is the one input that matters. 1.0 applies the LoRA at full strength; values below that blend it in more gently. Since it's a real LoRA merge (not a lightweight overlay), you can even go negative to push against the behavior, though nobody sane does that with outpainting LoRAs. The −100 to 100 range is the loader's generic allowance, not a suggestion.

Where it goes in the workflow

This LoRA isn't a general-purpose add-on - it's the outpaint mode of the LTX 2.3 unified sampler. If you've used the CRT unified sampler workflow, you've met the Models Pipe: it bundles the model, VAEs, and CLIP, and has optional slots for mode-specific models. This node's output feeds the model_outpaint input on LTX 2.3 US Models Pipe (CRT), which the unified sampler uses for V2V outpainting. That's the whole intended integration - load it, wire it into the pipe, outpaint.

Install

Standard pack 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. No extra per-node dependencies - it uses ComfyUI's core LoRA loading.

Gotchas

The first run downloads the LoRA (a few hundred MB, console progress), so the node looks like it's hanging on first execute. And read the tooltip on the pipe's latent_downscale_factor input if you're doing outpainting for real: the IC-LoRA family carries a latent-downscale factor in its metadata that the sampler needs to know about - the CRT nodes generally read it from the LoRA's metadata, so connect the loader output rather than hand-guessing the value. Also note this node loads the LoRA onto whatever model you pass it - pass the plain base model, not one you've already merged a different IC-LoRA into, or you'll stack behaviors you didn't ask for.