Beyond Scalar Rewards by Internalizing Reasoning into Score Distributions
Researchers have introduced Z-Reward, a novel teacher-student framework designed to improve text-to-image generation by better handling subjective visual preferences. The framework decouples complex reasoning from efficient reward deployment, with a large teacher model inferring score distributions and a smaller student model internalizing this reasoning for faster inference. This approach achieved high human preference accuracy and significantly improved text-to-image optimization performance compared to existing methods. AI
IMPACT Enhances AI image generation by providing more nuanced reward signals, potentially leading to higher quality and more preferred outputs.