search

RockeyCoss / SPO

Step-aware Preference Optimization: Aligning Preference with Denoising Performance at Each Step
https://arxiv.org/abs/2406.04314
157 stars 3 forks source link

About 1.5 #1

Closed ethanfel closed 5 months ago

ethanfel commented 5 months ago

Interesting, thanks for the XL model. Will you release a 1.5 model ? Our experiments with Stable Diffusion v1.5 and SDXL demonstrate that SPO significantly outperforms the latest Diffusion-DPO in aligning generated images with complex, detailed prompts and enhancing aesthetics, while also achieving more than 20× times faster in training efficiency.

liangzheng06 commented 5 months ago

Thanks for asking. We will release the SD-1.5 model shortly.