iamccs_ltx2_lora_stack (3 slots)

If you've stacked LoRAs in ComfyUI before, you know the pain: chaining three Load LoRA + Apply LoRA pairs across a graph, each adding a node hop and a place for the whole chain to break. This node collapses that into one box - three LoRA slots with independent strengths, one bundle out the other side. It's the LTX-2-specific member of the IAMCCS LoRA family, and it pairs with the pack's IAMCCS_ModelWithLoRA_LTX2 node, which applies the bundle to your model.

The important framing, because it differs from the pack's older WAN LoRA stack: this node does not do WAN-style key remapping. The original IAMCCS Wan stack was built to remap keys so WAN LoRAs load in native pipelines. The LTX-2 version is deliberately conservative - it standardizes only what's safe (converting legacy lora_down/lora_up keys to lora_A/lora_B) and otherwise assumes the LoRA and base model already match families. The author's own comment warns that aggressive remapping causes "widespread 'key not loaded' spam." Fair trade.

How it works

Each slot is a dropdown of every file in your models/loras folder plus "no":

• lora1 / strength1 (default 1.0), lora2 / strength2, lora3 / strength3.

Slots with "no" or strength 0 are skipped. The state dicts are loaded and lightly standardized, then bundled into a single LORA output. There's also an optional lora input of type LORA - if you already built a stack elsewhere (or chained another one of these), it appends those entries, so you can exceed three LoRAs by daisy-chaining stacks.

Output: one LORA - feed it to IAMCCS_ModelWithLoRA_LTX2 (or any node that takes a LORA bundle).

Install

ComfyUI Manager → "IAMCCS", or:

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

Restart. No model downloads from the pack - it reads whatever's already in your loras folder.

Gotchas

The honesty knob is the log output. After loading, it prints how many LoRAs and how many total parameter keys made it in. The apply side (the paired node) suppresses "lora key not loaded" lines for known mismatched prefixes - transformer_blocks. and text_embedding_projection. - because they spam, but suppression does not change which weights loaded. If your LoRA's effect is weak or absent, that's the smoke signal: the base model and the LoRA aren't the same family, and the pack says as much in its own log line. Stack strength is also not free - two 1.0 character LoRAs rarely behave like one 2.0; expect to back the strengths down when layering.