AI Glossary / LoRA

🏗️ Models and training

LoRA (Low-Rank Adaptation) is an efficient fine-tuning technique: instead of modifying billions of model parameters, you train a small "adapter" that changes its behavior. The result weighs megabytes instead of gigabytes and trains on a single GPU. LoRA became especially popular in image generation - the community creates thousands of adapters teaching the model specific styles, characters or products. In language models, it's a popular way to cheaply adapt an open model to your own needs.