📊 statistics

Bridging data-driven priors via the score function for posterior sampling -- Comparative review and experimental study

This paper unifies diverse data-driven priors for Bayesian inverse problems through their score functions, demonstrating via comparative experiments on image restoration tasks that this common framework enables efficient and versatile posterior sampling.

Original authors: Elhadji Cisse Faye, Mame Diarra Fall, Sylvain Delchini, Nicolas Dobigeon

Published 2026-06-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Elhadji Cisse Faye, Mame Diarra Fall, Sylvain Delchini, Nicolas Dobigeon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a giant jigsaw puzzle, but someone has thrown away half the pieces, smudged the picture with fog, and then scattered the remaining pieces across the floor. This is what scientists call an "inverse problem." In the real world, this happens when we try to fix blurry photos, fill in missing parts of an image, or see details that are too small for our cameras to catch.

The paper you're asking about is like a new, super-smart strategy for solving these puzzles. Here is how it works, broken down into simple ideas:

1. The Problem: Guessing the Missing Pieces

When we have a damaged image (the "observation"), we know the rules of how it got damaged (like a blurry lens or missing pixels). But to fix it, we need to guess what the original picture looked like.

Traditionally, scientists tried to fix this by writing down strict mathematical rules about what a "good" picture looks like (e.g., "edges should be sharp," "colors should be smooth"). But these hand-written rules are often too rigid. They miss the complex, messy beauty of real life, like the texture of a rock or the curve of a face.

2. The New Idea: Learning from a Library of Images

Instead of writing rules, the authors suggest using a "library" of thousands of real images (like faces or landscapes) to teach a computer what a "good" image looks like. The computer learns a Prior—a mental model of what reality usually looks like.

The paper focuses on a specific mathematical tool called the Score Function. Think of the Score Function as a compass.

If you are lost in a foggy forest (a damaged image), the compass doesn't tell you exactly where home is.
Instead, it points in the direction of "more likely to be home."
If you follow the compass step-by-step, you eventually find your way to a clear, sharp image.

3. The Old Way vs. The New Way

The paper compares two ways of following this compass:

The Old Way (ULA): Imagine trying to walk through the forest while holding a heavy, wobbly stick. If the terrain is very steep (a very blurry or damaged image), the stick gets stuck, and you move very slowly or get stuck in a loop. This is the "Unadjusted Langevin Algorithm" (ULA). It works, but it's slow and clumsy on tough problems.
The New Way (LwSGS): The authors propose a new method called Langevin-within-Split Gibbs Sampler (LwSGS).
- The Metaphor: Instead of trying to walk the whole path at once, imagine you have a helper.
- Step 1: You ask the helper to guess what the image should look like based on their memory (the "Prior" or the Score Function compass).
- Step 2: You take that guess and check it against the actual blurry photo you have.
- Step 3: You combine the two to make a better guess.
- The Magic: By splitting the problem into these two smaller steps, the "helper" can move much faster and more accurately through the forest, even when the terrain is very steep (severely damaged images).

4. What Did They Test?

The authors tested this new "compass-following" strategy on four different types of "memory libraries" (priors):

Denoising: A library that knows how to clean up noisy photos.
Score-Based Models: A library trained specifically to point toward "real" images.
Normalizing Flows: A library that learns how to stretch and twist simple shapes into complex images.
Convex Ridge Regularizers: A mathematical library that learns patterns from data.

They tested these on two main tasks:

Inpainting: Filling in missing holes in a picture (like removing a watermark or a scratch).
Super-Resolution: Turning a tiny, blurry pixelated image into a big, sharp one.

5. The Results

The paper found that their new method (LwSGS) was faster and more accurate than the old method (ULA) for all four types of libraries.

It produced sharper images with better details.
It didn't get "stuck" as easily.
It was also able to tell you how sure it was about its answer. For example, in the missing parts of a puzzle, it could say, "I'm pretty sure this is a tree, but I'm less sure about the leaves," showing a map of uncertainty.

6. Real-World Test: Rocks and Dirt

Finally, they didn't just use pretty pictures of faces. They used this method on geological samples (rocks and tailings) taken by the French Bureau of Geological and Mining Research.

The Problem: The photos of these rocks were blurry because the camera couldn't focus on the rough, uneven surfaces.
The Solution: They used the new method to sharpen the blurry rock photos.
The Outcome: The restored images were much clearer, allowing geologists to see the tiny particles and cracks that were previously invisible. This proves the method works not just on pretty pictures, but on messy, real-world data.

Summary

In short, this paper introduces a smarter, faster way to fix damaged images. Instead of struggling through the fog alone, it uses a "helper" system that splits the problem into manageable steps. It works with different types of "memory libraries" and has been proven to restore both digital photos and real-world images of rocks, making the invisible visible again.

Technical Summary: Bridging Data-Driven Priors via the Score Function for Posterior Sampling

Problem Statement
Image restoration (IR) tasks, such as denoising, deblurring, inpainting, and super-resolution, are typically formulated as linear inverse problems ($y = Ax + n$). Within a Bayesian framework, solving these ill-posed problems requires defining a prior distribution $p(x)$ to regularize the solution. While traditional model-based priors (e.g., Total Variation) are well-understood, they often fail to capture the sophisticated structures of natural images. Conversely, data-driven priors learned via deep learning (e.g., Generative Adversarial Networks, Variational Autoencoders, Normalizing Flows) offer superior expressiveness but pose significant challenges for posterior sampling.

A primary challenge in Bayesian IR is the high dimensionality of the image space and the non-smoothness of regularization functions, which hinder the efficiency of Markov chain Monte Carlo (MCMC) algorithms. Existing methods often rely on the Unadjusted Langevin Algorithm (ULA) to sample from the posterior. However, ULA can struggle with severely ill-conditioned problems due to step-size constraints linked to the Lipschitz constant of the likelihood function. Furthermore, many advanced data-driven priors (such as those based on denoisers or score-based generative models) are defined implicitly, making direct sampling difficult without a unified theoretical framework.

Methodology
The paper proposes a unified framework to sample from posterior distributions defined by diverse data-driven priors, provided they can be expressed via their score function ( $s(x) = \nabla_x \log p(x)$ ). The core methodology relies on the Langevin-within-Split Gibbs Sampler (LwSGS).

Asymptotically Exact Data Augmentation (AXDA): The algorithm introduces an auxiliary variable $z$ to decompose the posterior distribution into an augmented form:
$\pi_\rho(x, z) \propto \exp\left[ -f(x, y) - g(z) - \frac{1}{2\rho^2}\|x - z\|^2 \right]$
where $f(x,y)$ is the data fidelity term and $g(z)$ is the regularization potential. As the coupling parameter $\rho \to 0$ , the marginal distribution of $x$ converges to the target posterior.
Split Gibbs Sampling (SGS): The algorithm alternates between sampling $x$ and $z$ :
- Sampling $x$ : Given $z$ , the conditional distribution $p(x|y, z)$ is Gaussian (for linear Gaussian inverse problems), allowing for efficient, closed-form sampling. Crucially, this step decouples the step size from the Lipschitz constant of the likelihood, addressing a major limitation of standard ULA.
- Sampling $z$ : Given $x$ , the conditional distribution $p(z|x)$ is sampled using a Langevin Monte Carlo (LMC) step. This step explicitly utilizes the prior score function:
  $z^{(t+1)} = z^{(t)} + \gamma \left( s(z^{(t)}) + \frac{1}{\rho^2}(x^{(t)} - z^{(t)}) \right) + \sqrt{2\gamma}\epsilon^{(t)}$
  where $s(z) = \nabla \log p(z)$ is the score function of the prior.
Unified Prior Representation: The paper demonstrates that four distinct classes of data-driven priors can be unified under this score-function perspective:
- Regularization-by-Denoising (RED): The score function is derived directly from the denoising residual ( $s(x) \propto -(x - D(x))$ ).
- Score-Based Generative Models (SGM): The score function is explicitly learned via score matching (or denoising diffusion probabilistic models).
- Normalizing Flows (NF): The score function is computed via automatic differentiation of the log-density derived from the change of variables formula.
- Convex-Ridge Regularizers (CRR): The score function corresponds to a bias-free single-layer CNN output.

Key Contributions

Unification of Priors: The paper establishes that RED, SGM, NF, and CRR priors, despite their different origins and motivations, can all be integrated into a single sampling framework via their score functions.
Algorithmic Versatility: It shows that LwSGS is a generic algorithm capable of handling any differentiable log-prior with an explicit or efficiently computable score function, overcoming the limitations of simple ULA schemes.
Decoupling of Constraints: By splitting the problem, the method allows the sampling step for the prior to be independent of the data-fidelity term's conditioning, potentially improving mixing properties in ill-posed problems.

Experimental Results
The authors evaluated the LwSGS framework against ULA-based counterparts (PnP-ULA, NF-ULA, and newly designed SGM-ULA and CRR-ULA) on two tasks: image inpainting and single-image super-resolution.

Datasets: Experiments were conducted on 30 RGB images from the FFHQ and ImageNet datasets.
Performance Metrics: The study utilized PSNR, SSIM, LPIPS (perceptual similarity), and Integrated Autocorrelation Time (IAT) to measure reconstruction quality and sampling efficiency.
Findings:
- Reconstruction Quality: LwSGS consistently produced high-quality reconstructions that matched ground-truth details closely. In inpainting tasks, it effectively filled missing regions; in super-resolution, it preserved intricate details.
- Comparison with ULA: LwSGS generally outperformed ULA-based methods, particularly when using RED and SGM priors, yielding sharper details and higher fidelity. Even with NF and CRR priors, LwSGS showed competitive or superior performance.
- Efficiency: The LwSGS algorithm demonstrated faster convergence, indicated by lower IAT values and faster-decaying autocorrelation functions compared to ULA.
- Uncertainty Quantification: As a sampling method, LwSGS enabled the estimation of pixel-wise uncertainty (standard deviation), showing higher uncertainty in inpainted regions as expected.
- Real-World Application: The framework was applied to restore real-world geological tailing samples (deblurring out-of-focus images). Using a RED prior, the method successfully recovered sharp details from blurred images, demonstrating practical utility in geosciences.

Significance and Claims
The paper claims that the LwSGS framework provides a versatile and systematic alternative to ULA for Bayesian inverse problems involving score-function-based priors. Its significance lies in:

Broad Applicability: It is not limited to specific generative models but can handle any prior where the score function is accessible, bridging the gap between various data-driven regularization paradigms.
Practical Efficiency: The splitting strategy effectively mitigates the conditioning issues inherent in ULA, leading to better mixing and faster convergence in high-dimensional spaces.
Real-World Utility: The successful restoration of geological samples demonstrates that the method is robust enough for operational pipelines where acquisition protocols may be suboptimal.

The authors conclude that while the current work focuses on specific priors, the unified framework is sufficiently general to handle any posterior distribution defined by a broad class of score function-based priors, paving the way for future integration of diverse deep learning models into Bayesian inference.