LoRA Stack (Model In?Out) LTX-2

This is the "I don't want an extra node hop" version of the LTX-2 LoRA stack. Where the plain iamccs_ltx2_lora_stack outputs a LORA bundle that you then feed into a separate apply node, this one merges both steps: feed in your base model, pick up to three LoRAs with strengths, and out the other side comes a single patched MODEL, ready for the sampler. Model in, model out - hence the name.

If you're building the kind of graph this pack targets - long LTX-2 video pipelines where every extra node is another thing to break - collapsing stack+apply into one node is a genuine win for readability. It also keeps the model lineage obvious: the output model is exactly your base model with those three LoRAs baked in, nothing else.

How it works

• model - your base diffusion MODEL (LTX-2 or 2.3).
• lora1/strength1 (default 1.0), lora2/strength2, lora3/strength3 - dropdowns over your loras folder; "no" or strength 0 skips a slot.
• Optional lora (LORA) input - appends entries from an already-built stack, so you can chain past three.

Internally it loads each selected state dict, runs the same conservative key standardization as the rest of the LTX-2 LoRA family (legacy lora_down/lora_up → lora_A/lora_B, no aggressive WAN remapping), then applies them one at a time with comfy.sd.load_lora_for_models. Output: a patched MODEL.

If every slot is "no", it returns your input model unchanged with a warning - so it's safe to leave in a graph even before you've picked LoRAs.

Install

ComfyUI Manager → "IAMCCS", or:

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

Restart. No extra deps; reads your existing loras folder.

Gotchas

Same family-matching warning as the rest of these nodes: the apply step suppresses the noisy "lora key not loaded" lines for common mismatch prefixes, and the author's own log adds "if the LoRA effect is weak/absent, verify the base model matches the LoRA family." A LoRA trained for LTX-2 will partially load on LTX-2.3 and quietly under-deliver - that's the mismatch case. Also note there's no CLIP output here; it patches the model only. If your workflow needs the text encoder conditioned separately, keep that wiring independent. And watch the total stack strength: three maxed slots will fight each other before they help.