How it works
Training a full image model takes huge amounts of data and computing power. LoRA, short for low-rank adaptation, is a shortcut. The original model is frozen, and a small set of extra values is trained beside it. Those extra values nudge the model's output toward the new subject. Because only the small part is trained, the job fits on consumer hardware, and the result is a compact file that is easy to share.
To make one you collect a set of images of the subject, write a caption for each, and run a training tool for a number of steps. The settings matter. With too few steps the model does not learn the subject. With too many it memorizes the training pictures and cannot draw the subject in new poses, which is called overfitting. At generation time you load the LoRA with the base model and set its strength.
Creators meet LoRAs mainly through AI characters and brand styles. A character LoRA is one of the standard ways to keep the same face across hundreds of images. A style LoRA gives every image the same look. A LoRA only works with the model family it was trained on. Training on photos of a real person without their consent is a legal and ethical problem, not a technical detail.
A worked example
A creator has designed an AI character and wants the same face in every future image.
- Collect a varied set of images of the character: different angles, expressions, lighting and backgrounds.
- Caption each image and choose a rare trigger word that will call the character up in prompts.
- Train the LoRA, saving a copy at several points, then test each copy on the same prompts.
- Keep the copy that holds the face while still following new prompts. Later copies that redraw the training photos are overfitted.
An illustration, not a real account. Any numbers in it are made up to show the method, and are not benchmarks.
Common mistakes
- Using near-identical training images. Variety teaches the subject. Repetition teaches the photo.
- Training for too long. An overfitted LoRA copies its training images.
- Loading a LoRA with the wrong base model. It has to match the model family it was trained on.
- Training on a real person's photos without permission.
Questions people ask
How many images do I need to train a LoRA?
There is no single number. It depends on the subject, the model and the variety of the images. A smaller varied set usually beats a larger repetitive one. The guide linked on this page works through the trade-offs.
Is a LoRA the same as fine-tuning?
It is a lightweight form of fine-tuning. Full fine-tuning changes the whole model and produces a full-size file. A LoRA trains a small add-on and leaves the base model untouched.
Can I use more than one LoRA at once?
Yes, for example one for a character and one for a style. They can interfere with each other, so lower their strengths and test.
Sources
The paper that introduced LoRA, originally for language models. It keeps the pre-trained weights frozen and trains small added matrices.
arXiv (Edward J. Hu and others): LoRA: Low-Rank Adaptation of Large Language Models 17 June 2021
Hugging Face describes LoRA as a lightweight training technique that inserts a smaller number of new weights and trains only those, which makes training faster and produces smaller files.
Every link was opened on 13 October 2026. Products and policies change, so check the source before you rely on a figure.