LTX2.5 IC Upscale LoRA (CRT AutoDL)

There are two ways to make a video bigger: sharpen it after the fact, or teach the generator itself to add the detail. LTX2.5 IC Upscale LoRA (CRT AutoDL) is the second approach - it loads the LTX-2.3 22B upscale IC-LoRA and applies it to your model, so the upscale happens inside the generation as a conditioning task rather than as a dumb post-pass that can't invent anything new.

This is the difference that matters with latent upscalers and pixel upscalers alike: a conventional upscaler interpolates existing information, while an IC upscale LoRA gives the model a learned sense of what should exist at the higher resolution - texture, fine structure, the things a straight resize can only fake. In the LTX world, where the model already renders at half your requested resolution and upscales internally, that distinction is the difference between soft and crisp.

How it works

Standard CRT AutoDL LoRA mechanics: model in, strength_model (default 1.0) in, patched MODEL out. The node auto-downloads the adapter on first run and applies it at the given strength.

And like the union-control node, it also emits latent_downscale_factor - the reference downscale factor read from the LoRA's metadata, which the CRT loader surfaces as a float. The upscale path in LTX workflows has its own latent-scaling math, and having the correct factor handed to you (instead of tuned blind) is a genuinely useful touch. It defaults to 1.0 if metadata is missing, and it's still emitted at strength_model = 0 so you can read the factor without applying the LoRA.

Outputs: MODEL (back into your sampler/upscale pass) and latent_downscale_factor (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 a community repack on first run.

Common issues

• Subtle effect - IC upscale LoRAs are usually used in concert with the latent upscale stage, not as a standalone. If you expected a dramatic sharpening with no other pipeline changes, that's not the workflow it's built for.
• Forgetting the downscale factor - the resolution math matters. Wire latent_downscale_factor into your scaling, or results drift.
• Update breakage - red socket or NaN → right-click → Fix node (recreate).