CREPE🥞: Controlling diffusion with REPlica Exchange

CREPE🥞 is an inference-time control algorithm for diffusion models (both Gaussian diffusion and CTMC!).

Illustration of CREPE (right) and comparison with SMC control (left):

Example of CREPE for prompted reward-tilting on ImageNet-512:

This repo contains various applications of CREPE:

• Tempering on Boltzmann distribution of Alanine Di/Tetra/Hexa-peptide
• Trajectory stitching for maze
• Debiasing CFG for image generation
• Prompted reward-tilting for image generation
• Debiasing CFG on CTMC

The structure is as follows:

applications
    ├── tempering
    │        - code for inference-time tempering for Boltzmann sampling
    ├── stitching
    │        - code for stitching trajectories in the maze
    ├── cfg
    │        - debiasing cfg on ImageNet
    ├── reward-tilting
    │        - prompted reward-tilting on ImageNet
    └── ctmc
            - debiasing cfg for CTMC models 

The required environment may differ for each application. Please refer to their own folder.

Do not hesitate to reach out if you have any questions!

Reference

@article{he2025crepe,
  title={CREPE: Controlling Diffusion with Replica Exchange},
  author={He, Jiajun and Jeha, Paul and Potaptchik, Peter and Zhang, Leo and Hern{\'a}ndez-Lobato, Jos{\'e} Miguel and Du, Yuanqi and Syed, Saifuddin and Vargas, Francisco},
  journal={arXiv preprint arXiv:2509.23265},
  year={2025}
}