LoRA fine-tunes a model by training small low-rank adapter matrices that get added to the existing weights, rather than updating the full weight matrix. The adapter is typically less than 1% of the size of the full model.
LoRA dramatically lowers the cost of fine-tuning (a few hours on a single GPU instead of weeks on a cluster) and lets you swap personalities or styles by swapping adapters. It is the dominant fine-tuning technique in 2026 for open-source models. Stable Diffusion communities use LoRAs heavily for character and style control.