Skip to main content

What Is It: LoRA


LoRA stands for "Low Rank Adaptation"

In short, LoRA's are kinda like mods for a video game, with the game in this case being your SD Checkpoint.

LoRA's are trained on a variety of visual data specifically relating to what the LoRA is for. This typically will be specific characters, or visual styles.

On top of this, while just loading a LoRA is "okay", they typically are going to have trigger words. Discussing trigger words are where we start to get a bit into the weeds with how a LoRA works. The trigger words help to tell it when to kick in and emphasize things a bit more. Now conceptually this does sound similar to how Textual Embedding works, as new words are coming into play. But it's important to note that it doesn't understand these trigger words like how an embed might teach the AI new words entirely.

It's better to think of it as a signal. When utilizing a LoRA, the AI doesn't really know what the trigger words mean. But it does know to utilize the info from the LoRA when the trigger word is detected. LoRAs can also be setup to use trigger words that are valid prompting entries as well, i.e. "Red Eyes". The AI itself would understand what you meant when you told it "Red Eyes", but if a LoRA has that as a trigger word, it's utilizing the data from it rather than what the AI itself might actually know.

It's important to note that some parts of utilizing LoRA's are similar to the process of what one might find with Textual Embedding. Though they're not to be confused.