What does LoRA stand for?

Low-Rank Adaptation. It refers to the mathematical technique that makes the add-on file small rather than to anything about images specifically.

How big is a LoRA file?

Typically tens to a few hundred megabytes, against several gigabytes for the base model it modifies.

Do I need a LoRA to generate good images?

No. They exist to add something the base model does not already know — your product, your character, a specific style. For general images the base model is what you want.

Can I use more than one LoRA at once?

Yes, each with its own strength. Two strong style LoRAs tend to fight; lowering the weights usually resolves it better than removing one.

Can I train my own LoRA?

Yes, and it is the most accessible form of model customisation — a modest set of images and a consumer graphics card is generally enough. Quality depends far more on the consistency of the training images than on their number.

Do LoRAs work with any model?

No. A LoRA is trained against a specific base architecture and will not transfer to an unrelated one. A LoRA for one model family will not load into another.

Are LoRAs free to use commercially?

Not automatically. Each carries its own licence, and the base model licence applies as well. Some community LoRAs explicitly prohibit commercial use.

AI image glossary

What is a LoRA?

A small trained add-on that teaches a base model a specific style, character or object without retraining the whole model.

A LoRA — short for Low-Rank Adaptation — is a small file that modifies a large image model to produce one specific style, character, object or aesthetic. Instead of retraining a model that runs to several gigabytes, a LoRA trains a compact set of adjustments, typically a few dozen to a few hundred megabytes, that are applied on top of the base model at generation time.

A small brass adapter ring resting on a large camera lens on warm paper

The problem LoRAs solve

A base image model knows an enormous amount in general and nothing about your particular subject. It has never seen your product, your character, your illustration style or your face. Teaching it any of those by retraining the full model requires hardware and data that put it out of reach for almost everyone, and produces another multi-gigabyte file for every variation.

Why "low rank" makes it small

The insight is that the adjustment needed to specialise a model is far simpler than the model itself. Rather than learning a new value for every weight, a LoRA learns a compact factorisation — two small matrices whose product approximates the change — and that approximation turns out to be good enough. The result is a file two or three orders of magnitude smaller than the model it modifies, trainable on a single consumer graphics card in minutes to hours.

What people actually train them on

Four things dominate. A style, so that everything generated comes out in one illustrative or photographic register. A character, so a consistent person or creature can appear across many images. An object, typically a product that has to look exactly like itself. And a concept the base model handles poorly — a specific garment construction, an architectural detail, a niche piece of equipment.

Stacking, weights and where it goes wrong

Several LoRAs can be applied at once, each with its own strength, which is where most of the practical difficulty lives. Two style LoRAs at full strength fight each other and produce mush; a character LoRA at too high a weight bleeds its training background into every image. The usual fix is lowering weights rather than removing files — most stacking problems are volume problems.

The licensing question people skip

A LoRA is a derivative of both the base model and the images it was trained on, and neither of those is automatically yours to redistribute. Community LoRAs carry their own licences, some of which forbid commercial use, and a LoRA trained on a living artist's work to reproduce that artist's style raises a question no licence file settles. Check before you build a business on one.

LoRA at a glance

Stands for: Low-Rank Adaptation
File size: Tens to hundreds of MB
Base model size: Several GB
Loaded at once: Several, with weights
Teaches: A style, character or object
Portable across models: No

Related terms

What a LoRA sits next to

What is a checkpoint?The model it attaches to.

What is Stable Diffusion?Where the ecosystem lives.

What is FLUX.1?A smaller, newer adapter scene.

What is Nano Banana?Consistency without training.

all photo and art stylesStyle without a training step.

terms explainedEvery term, defined.

Put it to use

Where this appears

the main generatorThe main generator — text to finished image.Open browse real prompts120 images with the exact prompts that made them.Open Image to ImageStart from a picture you already have.Open AI Photo EditorChange part of a photo, keep the rest.Open AI Image UpscalerEnlarge to 4K without the smearing.Open enhance a photoExposure, noise and old-photo restoration.Open masked object removalMasked removal that rebuilds the background.Open Blur BackgroundDepth-aware, not a uniform filter.Open add colour to an old photoPlausible colour for black and white.Open generate a logoMarks that still read at 16 pixels.Open AI Tattoo GeneratorDesigns that survive being tattooed.Open generate illustrationIllustration and painting, not photographs.Open generate a headshotStudio-grade portraits without a studio.Open passport photo requirementsThe exact specification, and how to meet it.Open AI Interior DesignRestyle a room you can photograph.Open AI Image GlossarySeed, denoise, inpainting, explained.Open

LoRA — questions people ask

What does LoRA stand for?: Low-Rank Adaptation. It refers to the mathematical technique that makes the add-on file small rather than to anything about images specifically.
How big is a LoRA file?: Typically tens to a few hundred megabytes, against several gigabytes for the base model it modifies.
Do I need a LoRA to generate good images?: No. They exist to add something the base model does not already know — your product, your character, a specific style. For general images the base model is what you want.
Can I use more than one LoRA at once?: Yes, each with its own strength. Two strong style LoRAs tend to fight; lowering the weights usually resolves it better than removing one.
Can I train my own LoRA?: Yes, and it is the most accessible form of model customisation — a modest set of images and a consumer graphics card is generally enough. Quality depends far more on the consistency of the training images than on their number.
Do LoRAs work with any model?: No. A LoRA is trained against a specific base architecture and will not transfer to an unrelated one. A LoRA for one model family will not load into another.
Are LoRAs free to use commercially?: Not automatically. Each carries its own licence, and the base model licence applies as well. Some community LoRAs explicitly prohibit commercial use.

Related AI image terms

Start creating today

Generate your first image with the best AI image generator.

Turn a sentence into a finished, photorealistic image in seconds — then refine every detail. No setup, no Discord, no GPU.

Generate without training a LoRA Get Started

Join 4,200+ creators using LaFoto