Omini Kontext LoRA Unload

A small node with a single job: remove a LoRA adapter from a pipeline you've already loaded and hand the clean pipeline back. Omini Kontext LoRA Unload is the "undo the LoRA" step for when you want a with-LoRA and without-LoRA comparison, or when you loaded an adapter early in the chain and a later branch needs the plain model - without loading the whole 12B pipeline a second time.

How it works

It calls pipeline.delete_adapters([adapter_name]) on the diffusers pipeline and returns the same OMINI_KONTEXT_PIPELINE object, minus the adapter. The only required input besides the pipeline is adapter_name, which defaults to omini_kontext - that's the adapter name every node in this pack uses by default, so unless you renamed things in the LoRA Loader, the default is right. If the adapter doesn't exist, the node doesn't explode; it prints a warning to the console and carries on, which is friendlier than most ComfyUI nodes would be.

Why you'd use it

The realistic scenario is A/B testing. Chain loader → LoRA Loader → LoRA Unload, and you can tee up two generation branches - one that keeps the LoRA-loaded pipeline and one that gets the unloaded pipeline - and run them back to back against the same seed. Because the model never leaves VRAM, the comparison is fast and it's the honest way to see whether your character LoRA is actually doing the work or whether the base pipeline would've given you the same insertion.

The honest take

This is a thin wrapper around a diffusers method, and its usefulness tracks how much you actually A/B test. If you never compare, you'll never touch it - you can just as easily reload a clean pipeline from the loader node. But when you need it, it's exactly right, and it's the smallest node in the pack: two inputs, one output, no settings to get wrong. Same install as the rest of the pack - clone into custom_nodes, pip install -r requirements.txt, restart.

One thing to keep in mind while you're here: this unloads the LoRA from the in-memory pipeline, it does not unload the model itself. If your goal was actually freeing VRAM, you want the loader node deleted or bypassed, not this. This node is about adapter state, not memory.