Back to glossary

LoRA

A small, efficient fine-tune layered on top of a base generative model to teach it a specific style or subject.

A LoRA (Low-Rank Adaptation) is a small, efficient fine-tune layered on top of a base generative model to teach it a specific style, character, or subject — without retraining the whole model. Because a LoRA only adds a small set of new parameters, it's cheap to train and easy to swap in and out of a base diffusion model.

LoRAs are how creators get a consistent character or brand style out of an otherwise general-purpose model, rather than re-prompting and hoping for consistency each time.

See also: Diffusion Model, ControlNet.

Put these ideas into motion.

Download

Frequently asked questions

LoRA stands for Low-Rank Adaptation, a technique for fine-tuning a large generative model efficiently by training a small set of additional parameters instead of the entire model.

A LoRA is small — often tens of megabytes instead of gigabytes — quick to train, and can be swapped in and out of a base model, making it a practical way to teach a model a specific character, style, or product without retraining it from scratch.

Ready to tell your story?

Describe an idea and watch the agent animate it — on your Mac, in minutes.