Back when AI profile pictures were the hype, I helped drawanyone reduce the VRAM requirements for dreambooth: a few-shot finetuning method to update the model on what a "face" of our subject should look like, allowing us to then prompt images of it. Through a mix of quantization and improving the regularization step, I halved the VRAM requirements from 24GB to 12GB, meaning we could now use 3080s instead of 3090s for training. This represented a 35% cost reduction for my client and turned them profitable.