Apply LoRA to MODEL (LTX-2, quiet logs)

The LTX-2 sibling of the flagship LoRA applier

If you've used IAMCCS_ModelWithLoRA on Wan, this is the same idea pointed at LTX-2: take a LORA stack and apply it to a MODEL, output a patched MODEL. Identical input contract - model (MODEL) + lora (LORA) in, patched MODEL out. The differences are tuned for LTX-2 specifically, and the "quiet logs" in the name isn't a joke, it's the point.

What "quiet logs" means

LTX-2 is a 19B DiT with a high-compression spatiotemporal VAE, and Lightricks' LoRA ecosystem includes many specialized adapters (control LoRAs for pose/depth/canny, style LoRAs, speed/distill variants). Applying those stacks through ComfyUI's standard load_lora_for_models machinery produces a wall of lora key not loaded warning lines for keys the adapter simply doesn't carry. On a long video run that's hundreds of lines of console noise per apply.

This node installs a temporary log filter that suppresses those "key not loaded" messages during the apply, counts how many were suppressed, and prints one tidy summary line at the end instead. If the filter suppressed a lot from a particular prefix, it adds a note: "If the LoRA effect is weak/absent, verify the base model matches the LoRA family." That's a genuinely useful diagnostic - a suppression count of thousands usually means you've applied a Wan or SDXL LoRA to an LTX-2 model, which will never work.

Under the hood it does the straightforward thing: loops the stack entries, calls comfy.sd.load_lora_for_models per entry with its strength, and returns the patched model. The filter is added and removed around the apply, so it doesn't mute your other logging.

When to pick this over the plain version

Honestly, for most LTX-2 work you could use either node - they're siblings, and the input/output contract is identical. Pick this one when you're running in a loop or a long batch where console spam makes real errors hard to spot. The parent pack's native LoRA system (stack + apply) is what fixes LoRA loading in native WANAnimate-style flows, and this is the LTX-2 flavor of the same pattern, used with the pack's LTX-2 LoRA stacks.

Install is the pack install - Manager, search "IAMCCS", or:

cd ComfyUI/custom_nodes
git clone https://github.com/IAMCCS/IAMCCS-nodes.git

Then restart. Env floor is ComfyUI 0.3.x+, Python 3.12+, Torch 2.8+; no extra pip packages for this node.

The one trap to remember: a quiet log doesn't mean a working LoRA. If the effect looks weak, the node's own suppressed-count summary is your clue. A big number with a matching prefix = wrong model family, not a bug. The "quiet" is the feature, but it's also a chance to miss a silent failure - check the summary line once in a while.