New method speeds up AI image training by learning rewards mid-process

Original: Latent Reward Registers for Diffusion Preference Alignment

Source: arxiv.org ↗