LTX2.5 IC Cnet LoRA (CRT AutoDL)

ControlNet-style conditioning is one of the things that made video models actually directable, and LTX2.5 IC Cnet LoRA (CRT AutoDL) is how you get it in the LTX2.5 stack. "Cnet" is the control-net LoRA - specifically the LTX-2.3 22B IC-LoRA Union Control adapter, which is what Lightricks shipped to give LTX the kind of structural conditioning (edges, structure, layout) that the ControlNet ecosystem gives SDXL. This node downloads it on first run and applies it to your model.

The "IC" in IC-LoRA stands for in-context/image-conditioning LoRA, and it's a big part of how the LTX family does control and editing without a separate ControlNet stack: small loadable adapters that teach a behavior - control, outpainting, upscaling, VFX-style edits - by conditioning on a reference. The union-control variant is the general-purpose structural conditioner, the one you reach for when you want the generation to follow an edge map or a structural guide instead of just a prompt.

How it works

Like every AutoDL LoRA node, it takes your loaded model and a strength_model float (default 1.0, range −100 to 100) and applies the LoRA, returning a patched MODEL plus - here's the CRT twist - a second output: latent_downscale_factor.

That second output is the reason this node is nicer than a raw LoRA loader. LTX control workflows don't just apply the adapter; they also need to adjust how the latent is scaled to match the conditioning, and the correct factor is baked into the LoRA's metadata. The CRT loader reads reference_downscale_factor from the file and hands it to you as a float, so you can feed it straight into the resolution/scaling math of your workflow instead of guessing (or copying a magic number from a forum post). It defaults to 1.0 if the metadata is missing, and it's still emitted when strength_model is 0 - so you can read the factor without applying the LoRA at all.

Outputs: MODEL (wire back into your sampler) and latent_downscale_factor (wire into your latent-scaling node).

Installing it

Part of CRT-Nodes:

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. The LoRA downloads from the official Lightricks repo on first run - it's a few GB, so the console progress bar is your friend.

Common issues

• Conditioning has no effect - union control needs the right conditioning setup (edge/structural input) feeding the model, not just the LoRA applied. Check the rest of the graph, and confirm strength_model isn't 0.
• Ignore latent_downscale_factor at your peril - workflows that don't scale the latent to match the control produce soft or offset results. Wire that float where your workflow computes resolution.
• Wrong generation's LoRA - this is the LTX-2.3 22B adapter; it's paired with the 2.5 stack by the pack, but if you're on a different base it won't behave.
• Update breakage - red socket or NaN → right-click → Fix node (recreate).