LoRAs ("Low-Rank Adaptations") are like small modular chunks that get tacked onto a generative AI model to introduce additional concepts, fine tune some aspect of it, etc, in a way that's generic enough that they work within a given family of checkpoints and sometimes across it within the same base model. Like something derived from SDXL might work with any other SDXL LoRA but probably won't because of the changes the major finetune strains of it made when retraining it (Pony XL, a notorious porn checkpoint derived from SDXL that spawned an entire ecosystem of further finetunes, in particular completely fried something and is incompatible with everything else).

In this case they're throwing a fit about Flux Kontext (a new finetune of Flux that's focused on image to image edits I believe? That's what I've inferred anyways) LoRAs designed for image-to-image edits to remove clothes are being removed from every major hosting site because they violate site policy and/or the law.

Tbh your definition hits, though technically think of it as extra specialist training or a filter to generate more specific results, there's ones for LLMs and image AIs, probably audio ones too, but I don't deal with those much so idk. LoRAs were the secret to that Danny Devito staring in a live action Little mermaid spoof that was circulating a few months ago.

Simple explanation: It's a plugin you use for fine tuning.

If you want to generate an image that always has a specific anime character for example, you use a LoRA for that character.

Same can be for poses or other things.