Rethinking Visual Reasoning in Text-to-Image Reward Modeling
The resulting reward model improves text-to-image generation on challenging compositional prompts (GenEval) by an average of 4.0% and 5.5% respectively, outperforming HPSv3 which was trained on 6x more preference pairs.
Shi-Ye Su, Xiao-Han Wang, Ludwig Schmidt et al.
· 0 citations