DreamDPO: Aligning Text-to-3D Generation with Human Preferences
via Direct Preference Optimization


Preprint 2025

Zhenglin Zhou1, Xiaobo Xia2†, Fan Ma3, Hehe Fan1, Yi Yang1†, Tat-Seng Chua2

1 Zhejiang University, 2 National University of Singapore, 3 Yale University

Paper Code

Video


Method Overview


Framework of DraemDPO. DreamDPO first constructs pairwise examples, then compares their alignment with human preferences using reward or large multimodal models, and lastly optimizes the 3D presentation with a preference-driven loss function. The loss function pulls the win example closer and pushes the lose example away.

DreamDPO with Reward Model

DreamDPO with Large Multi-modal Model

Citation 


@article{zhou2025dreamdpo,
  title={DreamDPO: Aligning Text-to-3D Generation with Human Preferences via Direct Preference Optimization},
  author={Zhou, Zhenglin and Xia, Xiaobo and Ma, Fan and Fan, Hehe and Yang, Yi and Chua, Tat-Seng},
  journal={arXiv},
  year={2025}
}