Hongyang Du, Yunfei Xie, Junjie Ye, Jiawei Yang, Xiaoyan Cong, Haodong Zhang, Yongchao Huang, Haiyu Wu, Zongxia Li, Shihang Gui, Dawei Liu, Runhao Li, Jingcheng Ni, Chen Wei, Randall Balestriero, Yue Wang
FuseReg introduces a regularization technique that mitigates the reconstruction-generation gap in representation autoencoders by training over random subsets of encoder layers. This approach achieves higher PSNR and reduces unguided gFID without modifying the pretrained encoder. It also benefits generation and diffusion training.
The proposed method improves the flexibility and performance of representation autoencoders, which is crucial for various computer vision tasks. It also reduces the need for retraining and decoder specialization, making it a practical and efficient solution. This can lead to better image generation and reduced reconstruction errors.
Thank you to arXiv for use of its open access interoperability. arxiv.org. aibrief.ai is not affiliated with or endorsed by arXiv.