Using RL to directly optimize human preferences, this approach is quite clean and much more elegant than stacking classifiers.

ME News Report, April 19 (UTC+8), recently, researchers developed a simple and sample-efficient online reinforcement learning technique for trained image generation models. This technique is seen as a potential, steerable alternative to classifier-free guidance methods, with its driving signal being any scalar reward, including human preferences. More information is provided via a Twitter link. (Source: InFoQ)