LoRA
A LoRA (Low-Rank Adaptation) is a small add-on file that teaches a base model a specific style, character, object or concept - without retraining the entire model.
A full image model has billions of parameters and is enormous to train. A LoRA - short for Low-Rank Adaptation - is a clever shortcut: instead of changing the whole model, it trains a tiny set of additional weights that nudge the base model toward a new concept. The resulting file is small (often a few megabytes to a couple hundred) and loads on top of a checkpoint.
What LoRAs are used for
- A consistent character or person across many images.
- A specific art style (a particular illustrator, film look, or brand aesthetic).
- A product, object or outfit you want to reproduce reliably.
How it works
You apply a LoRA on top of a base model and usually trigger it with a keyword from training. Most tools let you set a LoRA weight to dial its effect up or down, and you can stack several at once - say a character LoRA plus a style LoRA. Because it only adds a thin layer, a LoRA trains fast and is easy to share, which is why they are everywhere in the open-source ecosystem.
Try it in the generator
Put lora to work right now - free daily generations, commercial license included.
Related terms
- CheckpointA checkpoint is a saved AI model file containing the full set of trained weights. It is the complete "brain" that generates images - swapping checkpoints changes the entire look and capability.
- Diffusion modelA diffusion model is the type of AI that powers most modern image generators. It learns to turn random noise into a coherent image by reversing a step-by-step noising process.
- ControlNetControlNet is an add-on that conditions a diffusion model on a reference structure - such as a pose skeleton, edge map or depth map - so you control composition precisely, not just with words.
- Style transferStyle transfer applies the visual style of one source - an image, artist or aesthetic - onto your own content, keeping your subject while changing its look.