arxivcs.CV2026-07-13
Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation
Runhui Huang, Qihui Zhang, Zhe Liu, Yu Gao, Jie Wu, Hengshuang Zhao
In this paper, we propose SpectraReward, a training-free reward function that turns pretrained MLLMs into off-the-shelf reward models for image-generation reinforcement learning. Instead of asking the MLLM to judge a generated image or answer decomposed verification questions, Sp…