Abstract:Diffusion models have recently shown considerable potential in solving Bayesian inverse problems when used as priors. However, sampling from the resulting denoising posterior distributions remains a challenge as it involves intractable terms. To tackle this issue, state-of-the-art approaches formulate the problem as that of sampling from a surrogate diffusion model targeting the posterior and decompose its scores into two terms: the prior score and an intractable guidance term. While the former is replaced by the pre-trained score of the considered diffusion model, the guidance term has to be estimated. In this paper, we propose a novel approach that utilises a decomposition of the transitions which, in contrast to previous methods, allows a trade-off between the complexity of the intractable guidance term and that of the prior transitions. We validate the proposed approach through extensive experiments on linear and nonlinear inverse problems, including challenging cases with latent diffusion models as priors, and demonstrate its effectiveness in reconstructing electrocardiogram (ECG) from partial measurements for accurate cardiac diagnosis.

What problem does this paper attempt to address?

### Problems Addressed by the Paper This paper aims to address the challenges of sampling from the posterior distribution when using Denoising Diffusion Models (DDM) as priors in Bayesian inverse problems. Specifically, traditional Markov Chain Monte Carlo (MCMC) methods cannot be directly applied due to the lack of an explicit and tractable probability density function for the DDM prior. Additionally, gradient-based MCMC methods are often inefficient and prone to getting stuck in local modes of the posterior distribution. ### Solution The authors propose a novel method called **Midpoint Guidance Posterior Sampling (MGPS)**. This method addresses the aforementioned issues by decomposing the denoising step into two intermediate stages: 1. **Denoising Step**: First, the sample \( X_{k+1} \) is denoised to an intermediate state \( X_{\ell_k} \), where \( \ell_k < k \). 2. **Noising Step**: Then, unconditional noising is applied to obtain \( X_k \). This decomposition introduces an additional degree of freedom, as it only requires estimating the guidance term at the intermediate step \( \ell_k \) rather than at step \( k+1 \). The MGPS algorithm leverages Gaussian variational approximation and guidance approximation at the intermediate step \( \ell_k \) to develop a principled denoising transition approximation method. ### Experimental Validation The authors extensively validated the effectiveness of MGPS through experiments on both linear and nonlinear inverse problems, such as image restoration, super-resolution, phase retrieval, deblurring, JPEG dequantization, and high dynamic range imaging. Notably, they also applied this method to the interpolation task of 12-lead electrocardiograms (ECG) and demonstrated its superior performance in these tasks. ### Main Contributions - **Innovative Denoising Step Decomposition**: By introducing an intermediate step \( \ell_k \), the complexity of the guidance term is reduced, improving sampling accuracy. - **Efficient Variational Approximation**: Utilizing Gaussian variational approximation and guidance approximation, the authors developed an efficient and principled denoising transition approximation method. - **Extensive Experimental Validation**: The method was validated on various linear and nonlinear inverse problems, showcasing MGPS's superior performance across different tasks. ### Conclusion MGPS provides an effective method to address the challenges of posterior sampling when using DDM priors in Bayesian inverse problems, particularly in cases of highly multimodal posterior distributions, strongly nonlinear measurement processes, or data heavily contaminated by noise.

Variational Diffusion Posterior Sampling with Midpoint Guidance

A Variational Perspective on Solving Inverse Problems with Diffusion Models

Amortized Posterior Sampling with Diffusion Prior Distillation

Diffusion Prior-Based Amortized Variational Inference for Noisy Inverse Problems

Sequential Posterior Sampling with Diffusion Models

Diffusion Posterior Sampling for General Noisy Inverse Problems

Diffusion State-Guided Projected Gradient for Inverse Problems

Diffusion posterior sampling for magnetic resonance imaging

Divide-and-Conquer Posterior Sampling for Denoising Diffusion Priors

Posterior Sampling in High Dimension via Diffusion Processes

Diffusion Model Based Posterior Sampling for Noisy Linear Inverse Problems

Variational Markov chain Monte Carlo for Bayesian smoothing of non-linear diffusions

Posterior sampling via Langevin dynamics based on generative priors

Enhancing Diffusion Posterior Sampling for Inverse Problems by Integrating Crafted Measurements

Denoising Diffusion Variational Inference: Diffusion Models as Expressive Variational Posteriors

Tweedie Moment Projected Diffusions For Inverse Problems

Score-Based Variational Inference for Inverse Problems

Conditional Variational Diffusion Models

Subspace Diffusion Posterior Sampling for Travel-Time Tomography

Removing Structured Noise with Diffusion Models