CREPE: Controlling Diffusion with Replica Exchange

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: He, Jiajun, Jeha, Paul, Potaptchik, Peter, Zhang, Leo, Hernández-Lobato, José Miguel, Du, Yuanqi, Syed, Saifuddin, Vargas, Francisco
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866908860853256192
author He, Jiajun
Jeha, Paul
Potaptchik, Peter
Zhang, Leo
Hernández-Lobato, José Miguel
Du, Yuanqi
Syed, Saifuddin
Vargas, Francisco
author_facet He, Jiajun
Jeha, Paul
Potaptchik, Peter
Zhang, Leo
Hernández-Lobato, José Miguel
Du, Yuanqi
Syed, Saifuddin
Vargas, Francisco
contents Inference-time control of diffusion models aims to steer model outputs to satisfy new constraints without retraining. Previous approaches have mostly relied on heuristic guidance or have been coupled with Sequential Monte Carlo (SMC) for bias correction. In this paper, we propose a flexible alternative based on replica exchange, an algorithm designed initially for sampling problems. We refer to this method as CREPE (Controlling with REPlica Exchange). Unlike SMC, CREPE: (1) generates particles sequentially, (2) maintains high diversity in the generated samples after a burn-in period, and (3) enables online refinement or early termination. We demonstrate its versatility across various tasks, including temperature annealing, reward-tilting, model composition and classifier-free guidance debiasing, with competitive performance compared to prior SMC methods.
format Preprint
id arxiv_https___arxiv_org_abs_2509_23265
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle CREPE: Controlling Diffusion with Replica Exchange
He, Jiajun
Jeha, Paul
Potaptchik, Peter
Zhang, Leo
Hernández-Lobato, José Miguel
Du, Yuanqi
Syed, Saifuddin
Vargas, Francisco
Machine Learning
Inference-time control of diffusion models aims to steer model outputs to satisfy new constraints without retraining. Previous approaches have mostly relied on heuristic guidance or have been coupled with Sequential Monte Carlo (SMC) for bias correction. In this paper, we propose a flexible alternative based on replica exchange, an algorithm designed initially for sampling problems. We refer to this method as CREPE (Controlling with REPlica Exchange). Unlike SMC, CREPE: (1) generates particles sequentially, (2) maintains high diversity in the generated samples after a burn-in period, and (3) enables online refinement or early termination. We demonstrate its versatility across various tasks, including temperature annealing, reward-tilting, model composition and classifier-free guidance debiasing, with competitive performance compared to prior SMC methods.
title CREPE: Controlling Diffusion with Replica Exchange
topic Machine Learning
url https://arxiv.org/abs/2509.23265