Conditional sampling within generative diffusion models

Zheng Zhao, Ziwei Luo, Jens Sjölund, Thomas B. Schön

Introduction

Generative diffusion models have revolutionized the field of data synthesis, particularly in image and language modeling. However, a significant open challenge remains: adapting these models for conditional sampling. In many real-world applications, such as medical imaging or climate modeling, we are not just interested in generating random samples from a distribution $\pi(x)$, but rather in sampling from a conditional distribution $\pi(x|y)$ given specific observations $y$. This is the core of Bayesian inverse problems, where the goal is to reconstruct an underlying signal from noisy or incomplete data.

Methodological Framework

The paper categorizes the current landscape of conditional sampling into three primary methodologies, each offering different trade-offs between computational cost and flexibility:

1. Joint Bridging and Doob’s Bridging

These methods incorporate the condition $y$ directly into the diffusion process. Doob’s bridging attempts to "pin" the stochastic process at both the observation and the target. To address numerical instabilities associated with exact pinning, Generalised Joint Bridging utilizes Schrödinger bridges to identify the most efficient transport path between distributions, reducing the variance of the estimator.

2. Forward-Backward Path Bridging (Filtering)

This approach frames the reverse diffusion process as a stochastic filtering problem. By treating the condition as a measurement taken during the generative trajectory, researchers can apply iterative refinement techniques, such as Gibbs filtering, to guide the sample generation toward the conditional target.

3. Feynman–Kac & Sequential Monte Carlo (SMC)

A "plug-and-play" alternative, this method leverages pre-trained unconditional models. It employs an ensemble of "particles" that are weighted and resampled based on a likelihood function $\pi(y|x)$. This effectively "steers" the pre-trained model toward the desired conditional outcome without requiring expensive re-training, though it comes at a higher inference-time computational cost.

Pedagogical Evaluation

To compare these methods, the authors utilize a 2D Gaussian Mixture Model (GMM) with a non-linear likelihood. This synthetic setup tests the algorithms' ability to handle multi-modal distributions and "extreme" conditions where the observation $y$ is statistically unlikely.

Results indicate that while both Conditional Dynamic Schrödinger Bridges (CDSB) and Feynman–Kac SMC can recover the general structure of the true conditional distribution, they exhibit minor biases compared to gold-standard Hamiltonian Monte Carlo (HMC) baselines. Notably, joint-data methods like CDSB struggle when the condition $y$ falls into regions with low training data density, highlighting a sensitivity to the underlying data distribution.

Key Insights and Challenges

A conceptual highlight of the review is the connection between Lagrangian and Eulerian perspectives. The authors note that conditional SDEs (which track individual paths) and Feynman–Kac models (which track the evolution of the whole distribution) are essentially two views of the same probability flow.

However, several hurdles remain for the field:

  • Computational Cost: Training-free methods like SMC are highly flexible but require maintaining large ensembles of samples to represent the distribution accurately.
  • Outlier Robustness: Generative models often fail or introduce significant bias when faced with "outlier" conditions that deviate significantly from the training set.
  • Convergence and Dimensionality: Unlike traditional MCMC, there is currently no principled way to diagnose if a generative diffusion sampler has converged. Furthermore, score estimation errors can accumulate in high-dimensional spaces, potentially leading to "hallucinations."

Conclusion

While generative diffusion models offer a promising path for high-dimensional conditional sampling, their reliability in solving complex inverse problems depends on overcoming biases in extreme conditions and improving computational efficiency. Future research must focus on robust score estimation and the development of formal convergence metrics to ensure these models can be safely deployed in critical scientific and medical domains.

Summary updated

Abstract

Generative diffusions are a powerful class of Monte Carlo samplers that leverage bridging Markov processes to approximate complex, high-dimensional distributions, such as those found in image processing and language models. Despite their success in these domains, an important open challenge remains: extending these techniques to sample from conditional distributions, as required in, for example, Bayesian inverse problems. In this paper, we present a comprehensive review of existing computational approaches to conditional sampling within generative diffusion models. Specifically, we highlight key methodologies that either utilise the joint distribution, or rely on (pre-trained) marginal distributions with explicit likelihoods, to construct conditional generative samplers.
BibTeX
@article{zhao2025conditional,
  author = {Zheng Zhao and Ziwei Luo and Jens Sjölund and Thomas B. Schön},
  title = {{Conditional sampling within generative diffusion models}},
  journal = {Philosophical Transactions of the Royal Society A: Mathematical, Physical and Engineering Sciences},
  year = {2025},
  url = {https://openreview.net/forum?id=C2Y3lMPtlP},
}

Loading related content...

Figure