LTX2.3 IC Upscale LoRA (CRT AutoDL)

LTX's speed is its superpower, and its softness at lower resolutions is the tax that comes with it - the community has said since the 13B days that you render LTX at a higher resolution to get the detail other models give you at 720p. LTX 2.3 fixed a lot of that with the rebuilt VAE, but upscaling is still a normal stage in a serious workflow. This node loads the IC-LoRA that exists specifically for that pass.

Specifically, it auto-downloads and applies Zlikwid's ltx2.3_upscale_ic-lora_06250 - a community-trained InContext LoRA for the LTX 2.3 latent upscaler. InContext LoRAs are LTX's control-adapter format (Lightricks' answer to ControlNet for this model family), and this one conditions the upscale so that going from the half-res stage to full resolution doesn't just enlarge - it reconstructs detail.

How it works

Mechanically it's the same pattern as the other AutoDL LoRA loaders: feed in a MODEL, it checks whether the file already exists in models/loras/, downloads it from HuggingFace with a progress readout if not, then applies it at your chosen strength_model (default 1.0, range −100 to 100) and hands back the patched MODEL.

Where it sits in a workflow: the LTX 2.3 two-stage path generates at half resolution, runs the spatial latent upscaler, and then refines. This LoRA is the "detail recovery" partner to that upscale - applied to the model so the refinement pass reconstructs the fine structure instead of blurring the upscale. If you're using the pack's LTX 2.3 Unified Sampler, it slots into the models pipe; if you're hand-rolling with the vanilla ComfyUI LTX nodes, apply it to the model before the refinement sampler.

Installing it

Same pack install as everything else: ComfyUI Manager → CRT-Nodes, or:

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

Restart, drop the node in, and the ~1GB download happens on first execution.

Where people get burned

This node applies a LoRA to a model - it does not upscale anything by itself. If you wire it in and expect sharper output without an upscaling stage, you'll get nothing. It has to run in combination with the latent upscaler (which the pack's AutoDL set also covers via the LTX 2.3 spatial upscaler) and a refinement pass. Think of it as the seasoning, not the meal. Also worth knowing: strength defaults to 1.0, but community IC-LoRA practice often lands lower - if you get oversaturated or "baked" detail, back strength off rather than assuming the node is broken.