Stable Diffusion Masterclass - Theory, Code & Application

Creating & Training LoRA Models - Fine-tuning Stable Diffusion

We've learned how we can use textual inversion to create and train new embeddings that allow us to generate images of new subjects, concepts, and styles. Now we'll learn about a new technique, called LoRA, that allows us to achieve the same goal as textual inversion, but through a different mechanism.

LoRA stands for Low Rank Adaption, and it's a technique that allows us to fine-tune Stable Diffusion models to generate images of new concepts and styles. By using this technique, we freeze the weights of a pre-trained Stable Diffusion model and inject new, trainable layers.