Abstract:Given an unconditional diffusion model $\pi(x, y)$, using it to perform conditional simulation $\pi(x \mid y)$ is still largely an open question and is typically achieved by learning conditional drifts to the denoising SDE after the fact. In this work, we express conditional simulation as an inference problem on an augmented space corresponding to a partial SDE bridge. This perspective allows us to implement efficient and principled particle Gibbs and pseudo-marginal samplers marginally targeting the conditional distribution $\pi(x \mid y)$. Contrary to existing methodology, our methods do not introduce any additional approximation to the unconditional diffusion model aside from the Monte Carlo error. We showcase the benefits and drawbacks of our approach on a series of synthetic and real data examples.

What problem does this paper attempt to address?

This paper discusses the challenges of conducting conditional simulations in diffusion models, particularly how to generate samples from the conditional distribution π(x|y) given an unconditional diffusion model π(x, y). Current methods typically achieve this goal by learning the conditional drift of a noise SDE, but this introduces additional approximations. The paper proposes a new Markov chain Monte Carlo (MCMC) method called Forward-Backward Bridging (FBB), which treats conditional simulation as an inference problem on a joint space involving partial SDE bridges. This method is based on performing forward simulation first and then "bridging" back to the target distribution using a backward particle filter. The advantage of the FBB method is that it does not require additional approximations for the unconditional diffusion model, aside from Monte Carlo error. The paper demonstrates the advantages and limitations of this method on a range of synthetic and real data, particularly in the areas of image analysis, protein folding, and statistical inference. Furthermore, when the noise process is manageable and separable, FBB can run at zero memory cost, providing an unbiased generalization to the work of Trippe et al. The main contributions of this paper include: 1. Introducing FBB, a novel particle Gibbs method for conditional simulation in diffusion models. 2. Demonstrating how to run a memoryless version of FBB at zero memory cost when the noise process is separable. 3. Showing improvements in sample quality of the method in high-dimensional benchmark tests. In the experimental section, the paper compares the proposed Gibbs-CSMC and PMCMC methods with other relevant methods such as standard particle filtering, distorted particle filtering, and conditional score matching, proving their advantages in terms of convergence, sample quality, and computational efficiency.

Conditioning diffusion models by explicit forward-backward bridging

Conditional Simulation Using Diffusion Schrödinger Bridges

Conditional Diffusion with Less Explicit Guidance via Model Predictive Control

On conditional diffusion models for PDE simulations

Conditional sampling within generative diffusion models

Diffusion Bridge Implicit Models

A Flexible Diffusion Model

Diffusion models as plug-and-play priors

Transport with Support: Data-Conditional Diffusion Bridges

Simple simulation of diffusion bridges with application to likelihood inference for diffusions

Schrödinger bridge based deep conditional generative learning

Diffusion with Forward Models: Solving Stochastic Inverse Problems Without Direct Supervision

Simulation of conditioned diffusions

Discrete Modeling via Boundary Conditional Diffusion Processes

A Training-Free Conditional Diffusion Model for Learning Stochastic Dynamical Systems

Non-Denoising Forward-Time Diffusions

Unveil Conditional Diffusion Models with Classifier-free Guidance: A Sharp Statistical Theory

Practical and Asymptotically Exact Conditional Sampling in Diffusion Models

ShiftDDPMs: Exploring Conditional Diffusion Models by Shifting Diffusion Trajectories