SDXL LoRA Training Script Released — Pivotal Tuning & Prodigy

AI Impact Summary

The community has developed a new training script for Stable Diffusion XL (SDXL) LoRAs using techniques like Pivotal Tuning and adaptive optimizers (Prodigy). This script combines optimizations from various existing LoRA training methods, including those used in Replicate's SDXL Cog trainer and Kohya's sd-scripts, to achieve high-quality LoRAs with minimal compute. The script focuses on efficient training through techniques like textual inversion and adaptive learning rates for the text encoder and UNet, offering a streamlined approach for users to create custom SDXL LoRAs.