As the issues: #4368 #4377 #4360 point out training LoRA currently requires a significant amount of memory. It's quite tough to get SDXL to use less memory given that the unet weights alone are a...