Inference-time Scaling of Diffusion Models through Classical Search

Overview

This repository provides the implementation of Inference-time Scaling of Diffusion Models through Classical Search. The approach leverages classical search algorithms to scale inference compute in diffusion models, improving efficiency and output quality.

Implementation

The imagenet folder provides the implementation of BFS and double-verifier for class-conditional image generation, the locomotion folder provides the Q-verifier test-time search for offline RL tasks, the text_to_image folder contains the BFS ablations, and the pointmaze folder provides the implementation of the long-horizon planning task. For installation of each task, refer to the instructions in each subfolder.

Citation

If you use this code, please cite:

@misc{zhang2025inferencetimescalingdiffusionmodels,
      title={Inference-time Scaling of Diffusion Models through Classical Search}, 
      author={Xiangcheng Zhang and Haowei Lin and Haotian Ye and James Zou and Jianzhu Ma and Yitao Liang and Yilun Du},
      year={2025},
      eprint={2505.23614},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2505.23614}, 
}

License

This project is licensed under the MIT License.