Small patch, big change.
LoRA is a technique that fine-tunes large AI models efficiently by training only small, low-rank matrices instead of the entire model.
Imagine you have a giant cookbook with thousands of recipes. To add your own twist, instead of rewriting the whole book, you just write a few sticky notes with your changes. LoRA does the same for AI models—it makes tiny, focused adjustments without redoing everything.
LoRA lets you customize powerful AI models (like image generators or language models) on a normal computer, saving time and money. You encounter it in tools that let you create specific styles or characters, like Stable Diffusion models for art.
Many think LoRA creates new AI models from scratch, but it actually only tweaks a small part of an existing one. It doesn't replace the base model; it adds a lightweight adapter that can be swapped or removed.
Low-Rank Adaptation (LoRA) is a parameter-efficient fine-tuning method that decomposes weight updates into low-rank matrices, which are trained while the original model weights remain frozen. This reduces the number of trainable parameters by orders of magnitude, enabling efficient adaptation to specific tasks or domains without catastrophic forgetting.