2023.12.20 Evening HelloWorld 3.2 AutoDPO Version Update

Well, the academic world has been buzzing lately... This afternoon, a team from Peking University has released a new method and model based on Diffusion-DPO improvement called AI-Self-Training-DPO-SDXL. As an early adopter of technology, I have also extracted lora from this new fine-tuned model and further optimized the integration with HelloWorld 3.1 version. This update seems to have some mystical improvements. I have generated multiple comparison images, and the new version appears to have a lower probability of finger and pupil breakdown.

If all goes well in the future, I plan to use this new method for further model training, hoping to achieve better results than the dreambooth training.