Accelerating Bayesian Posterior Sampling in Diffusion Models: A Review of Offline-Online Trade-offs

As expressive learned priors, diffusion models have attracted much attention in Bayesian inverse problems. However, when used directly for inversion, they face a core challenge: the score function of the noised posterior cannot be obtained from the prior score alone. This review systematically surveys the main methods for addressing this challenge and, for the first time, organizes the literature around the thread of "the cost split between offline preparation and online inference," providing researchers and engineers with a clear roadmap for acceleration.

Core Problem: The Posterior Score Is Unavailable

Bayesian inverse problems require both an informative prior model and computationally feasible posterior inference. Score-based diffusion models provide expressive learned priors; however, when used for inversion they introduce a conditioning problem: the score function of the noised posterior generally cannot be obtained from the prior score alone. This means that unconditional samplers cannot simply be applied as-is; additional mechanisms must be introduced to incorporate the observation information.

The review points out that this conditioning problem has spawned several technical routes, which sacrifice offline computation or online efficiency to varying degrees in exchange for feasible posterior sampling. Understanding these trade-offs is a prerequisite for selecting or designing acceleration schemes.

Score function is a core concept in diffusion models, referring to the gradient of the log probability density with respect to the data, i.e., (\nabla_x \log p(x)). In score-based generative models, the model learns the prior score (\nabla_x \log p(x)) and samples via annealed Langevin dynamics or reverse stochastic differential equations. However, in Bayesian inverse problems, what we actually need is the posterior score (\nabla_x \log p(x|y)), where (y) is the observation. By Bayes' theorem, (p(x|y) \propto p(y|x)p(x)), and taking the log and then the gradient gives (\nabla_x \log p(x|y) = \nabla_x \log p(x) + \nabla_x \log p(y|x)). Thus, the posterior score equals the prior score plus the likelihood score (\nabla_x \log p(y|x)). The key point is that the likelihood score is usually difficult to compute analytically or approximate, because it involves the observation model (p(y|x)), which may be nonlinear, high-dimensional, or have an unavailable gradient. This is the fundamental reason why "the posterior score cannot be obtained from the prior score alone," and it is the mathematical essence of the conditioning problem.

Method Landscape: Four Mainstream Routes

The review groups the main methods into four categories: likelihood guidance, spectral and projection methods, sequential Monte Carlo correction, and amortized conditional training. Each category handles the conditioning problem differently, and thus each has its own emphasis in terms of assumptions, advantages, and limitations.

• Likelihood guidance: Corrects the score using the observation likelihood during sampling, usually requiring approximation or gradient computation.
• Spectral and projection methods: Reduce the difficulty of conditioning through low-rank approximations or projection operations, potentially sacrificing some accuracy.
• Sequential Monte Carlo correction: Introduces mechanisms such as particle filtering to correct bias, improving asymptotic accuracy but increasing the online burden.
• Amortized conditional training: Trains a model conditioned on observations offline, front-loading the cost; online sampling is extremely fast but depends on the training distribution.

This classification is not a simple enumeration but reveals where different methods sit on the "offline-online" spectrum.

Sequential Monte Carlo (SMC) is a class of sampling methods based on particle filtering, suitable for high-dimensional state spaces and complex posterior distributions. In posterior sampling with diffusion models, SMC treats each denoising step of the diffusion process as an intermediate distribution and maintains a set of weighted particles to approximate these distributions. Specifically, at each step, particles are propagated according to the prior score, then reweighted according to the observation likelihood, and resampling is used to avoid weight degeneracy. The advantage of SMC correction is that it provides asymptotically consistent estimates and can theoretically correct the bias introduced by conditioning approximations. However, its computational cost grows linearly with the number of particles, and in high-dimensional problems it may face particle degeneracy (i.e., a few particles carrying most of the weight) and excessive computational burden. In addition, the design of the resampling strategy significantly affects performance and requires careful tuning.

The Offline-Online Perspective: How the Cost Is Split

The core contribution of the review is to organize methods by cost split. The offline preparation stage may include training conditional models, computing spectral bases, or precomputing caches; the online inference stage involves sampling iterations, likelihood evaluations, or particle updates. The amount of computation allocated to these two stages varies significantly across methods.

For example, amortized conditional training shifts most of the computation to the offline stage, requiring only a few sampling steps online, making it suitable for repeated inference scenarios; sequential Monte Carlo correction, by contrast, places more computation in the online stage to cope with unknown or changing observations. Likelihood guidance and spectral methods often lie in between, balancing online efficiency and offline preparation through approximation.

This perspective helps practitioners choose appropriate methods according to the application scenario (e.g., real-time requirements, whether the observation distribution is fixed) rather than blindly pursuing a single metric.

Assumptions and Limitations of Acceleration Strategies

The review emphasizes that every acceleration strategy rests on specific assumptions. Likelihood guidance may rely on approximations of the likelihood function or the availability of gradients; spectral methods assume that the posterior can be captured by low-dimensional structure; sequential Monte Carlo requires a sufficient number of particles and a resampling strategy; and amortized training requires offline data to cover the distribution of online observations.

While these assumptions bring acceleration, they also limit the applicability of the methods. For example, amortized conditional training may perform poorly on out-of-distribution observations, while sequential Monte Carlo faces the risk of particle degeneracy in high-dimensional problems. The review's discussion of these limitations points out directions for improvement in future research.

Open Problems and Future Directions

The article concludes by collecting several ongoing directions and open problems. Although the review does not elaborate on details, it can be inferred that how to adaptively allocate offline and online computation, how to combine the advantages of multiple methods, and how to design scalable acceleration schemes for high-dimensional complex inverse problems remain active research areas.

For practitioners, this review provides a structured decision framework: given a computational budget and problem characteristics, weigh the costs of offline preparation and online inference, and select or combine appropriate methods. As diffusion models are widely applied in scientific computing and image reconstruction, this systematic acceleration perspective will become increasingly important.

(This article is based on a review abstract from an rss source and does not introduce specific data not mentioned in the original text.)

One-line Summary

A review that organizes acceleration methods for diffusion models in Bayesian inverse problems by offline-online cost split.

Paragraph Summary

Addressing the conditioning challenge that the posterior score is unavailable for diffusion models in Bayesian inverse problems, this review systematically surveys four categories of methods: likelihood guidance, spectral and projection methods, sequential Monte Carlo correction, and amortized conditional training. Its core contribution is to use "the cost split between offline preparation and online inference" as an organizing thread, revealing the trade-offs in computational allocation across methods. For example, amortized conditional training front-loads cost to the offline stage, making online sampling extremely fast but dependent on training distribution coverage; sequential Monte Carlo correction places computation in the online stage, improving accuracy but increasing burden; likelihood guidance and spectral methods lie in between. The review also discusses the assumptions and limitations of each method, such as out-of-distribution generalization issues in amortized training and particle degeneracy risks in sequential Monte Carlo, and points out that adaptive computation allocation and hybrid method design are future directions. This work provides a structured decision framework for researchers and engineers.