Papers12:03Chinese content扩散模型偏好对齐新方法 Latent Reward Registers这篇论文给扩散模型加了个奖励寄存器,不用改生成器就能获得稠密奖励,训练比在线 RL 快 33 倍,做生成对齐的可以看看。#Latent Reward Registers#Diffusion Transformer#RG-OPD#diffusion modelsaarXiv cs.LG@Yuanshen Guan et al. (4)SourceRead laterReadWorth followingUsefulFollow Latent Reward Registers