A LoRA is a small add-on file (usually 50-200 MB) that teaches a base AI image model a specific style, character or visual concept without retraining the whole model. The base model stays frozen; the LoRA injects a small number of trained parameters during generation. LoRAs are how the Stable Diffusion ecosystem on CivitAI lets you reproduce a specific artist style, anime aesthetic or fictional character on demand. Training your own LoRA takes 20-60 minutes on a consumer GPU with 20-50 reference images.