Instant Reference LoRA Apply

This is the boring half of the pack, and that's the point. It takes a model, a clip, and a lora_stack in, applies every LoRA in the stack in order, and hands back patched model, clip, and the same lora_stack untouched. Nothing trains, nothing downloads, nothing is clever. It's the "apply" side of the train/apply split that Instant Reference LoRA Train sets up.

Why would you want a separate apply node instead of just using the combined Instant Reference LoRA? Flexibility. Training a LoRA on one checkpoint and applying it to a different one is a classic LoRA workflow move - and it's the one genuinely useful trick the split enables. You also get to apply at whatever strength you want, because the strengths in a stack entry were baked in when the stack was created.

Inputs

• model and clip - the checkpoint you actually want to generate with, which does not have to be the one the LoRA was trained on (LoRAs are designed to transfer, and a quick 50-step reference LoRA will happily ride on a related checkpoint).
• lora_stack - the stack output from Instant Reference LoRA Train, Instant Reference LoRA, or Instant Reference LoRA Load.

Outputs

• model / clip - patched and ready for your sampler.
• lora_stack - passed through unchanged, so you can keep chaining further apply nodes or inspect it downstream.

That pass-through is the subtle part: the stack is a description of LoRAs plus strengths, not a consumed resource. You can apply the same stack to several models in parallel, and each gets the full set of LoRAs.

Install

Same as the rest of the pack - ComfyUI Manager, search "Instant Reference", install, restart; or clone it manually:

cd ComfyUI/custom_nodes
git clone https://github.com/cstria0106/comfyui-instant-reference

Notes

The heavy machinery - the sd-scripts clone, the venv, the WD14 tagger download - is all on the training side, so this node is instant in a way the pack's name wishes everything were. It shares the pack's rough edges only insofar as it applies LoRAs through the same in-node loader, so if a trained LoRA was saved in an unusual format you'd hit that here too. For normal output from this pack, it just works.