| Last Updated | Version | Changes |
|---|---|---|
| 8/3/2023 | 1.0 | First version! This document will be kept up to date with SDXL 1.0 Training Developments |
<aside> ⚠️ This guide will provide; • The basics required to get started with SDXL training • Sample settings which produce great results • Tips gleaned from our own training experiences
This guide is not; • A full, comprehensive, LoRA training tutorial • A full Google Colab Tutorial (but we’ve covered the essentials!) • A Kohya Trainer Tutorial (but we’ve covered the essentials!) • Definitive settings which will always work 100% of the time - experimentation is required!
</aside>
SDXL 1.0 is a groundbreaking new model from Stability AI, with a base image size of 1024x1024 - providing a huge leap in image quality/fidelity over both SD 1.5’s 512x512 and SD 2.1’s 768x768. The model also contains new Clip encoders, and a whole host of other architecture changes, which have real implications for inference (generating images) and training.
Read our SDXL Overview Guide here for more information on getting started with SDXL.
Before SDXL launched, Stability AI teased us with posts detailing how easy it is to train against SDXL. There’s some truth to that - it is very forgiving, and many different settings produce extremely good results, however - the hardware requirements are higher than expected!
Training SDXL has significantly higher hardware requirements than training SD 1.5 LoRA. The community is still working out the best settings, and it will take some time for the training applications to be optimized for SDXL, but at time of writing (8/3/2023) we can safely say that;