Glossary · Updated Sep 2, 2026

LoRA (low-rank adaptation)

An efficient fine-tuning method that trains small adapter weights instead of the whole model, popular for custom image styles.

Definition

LoRA adds low-rank matrices to a model's layers and trains only those, cutting compute and storage dramatically. In image generation communities, LoRA files capture styles, characters, or objects and can be combined at inference time. The same idea applies to language models for domain adaptation.

Why it matters when choosing a tool

Many image and 3D tools offer custom style training that is LoRA under the hood, so understanding it explains why a few dozen reference images can produce a consistent look.

Where you will meet it

AI Image & Design, AI 3D & Game Assets

Related terms

Fine-tuning · Diffusion model · Open-weight model

Tools where this matters

Reviewed products in the related categories