DiT-Reward: Generative Representations for Text-to-Image Reward Modeling

DiT-Reward is a new model that repurposes a pretrained text-to-image Diffusion Transformer for the task of reward prediction in image generation. It demonstrates superior performance over HPSv3 on preference benchmarks, achieving 85.6% on HPDv2 and 77.6% on HPDv3, and shows that a lightweight learned head can still yield meaningful predictions when the generative backbone is frozen. The findings indicate that pretrained generative models can effectively enhance reward modeling and policy optimization, offering a 1.65x inference speedup while maintaining comparable memory usage.