Definition
LoRA (Low-Rank Adaptation) is a parameter-efficient fine-tuning method: the original model weights stay frozen, and only small low-rank matrices injected alongside selected layers are trained, adapting the model to a task or style with a fraction of the trainable parameters.
How it works
LoRA approximates a weight update as the product of two small matrices, trains only that bypass during fine-tuning, and merges it back into the base weights afterwards. Compared with full fine-tuning this cuts training cost and memory use, and the resulting artifacts are small.
Usage
LoRA is the standard way to customize large language models or diffusion models on limited hardware — teaching a text-to-image model a particular art style, or adapting a chat model to a vertical domain. Community fine-tunes are commonly shared as LoRA adapters.
Related and confusable terms
Diffusion models are often fine-tuned with LoRA: one is the model, the other the adaptation technique. LoRA is also distinct from quantization — quantization shrinks numerical precision, while LoRA decides what gets trained on top.