Diffusion Posterior Proximal Sampling (DPPS)

• Diffusion Posterior Proximal Sampling (DPPS) is a framework that combines diffusion-based generative priors with proximal updates to enforce measurement consistency in Bayesian inverse problems.
• It alternates between diffusion denoising steps and MAP-like proximal updates, balancing prior information with data-fidelity for robust posterior sampling.
• Empirical results demonstrate improvements in PSNR, SSIM, and reduced sample variance, supported by theoretical guarantees across linear and nonlinear observation models.

Diffusion Posterior Proximal Sampling (DPPS) is a principled family of algorithms for sampling from Bayesian posteriors in high-dimensional inverse problems using pretrained diffusion models as expressive priors. DPPS leverages the structure of posterior distributions—combining a diffusion-based prior with a (potentially nonlinear) likelihood—by alternating between prior-guided diffusion steps and proximal (measurement-consistent) updates, thus enforcing data fidelity without requiring retraining of the underlying generative model. Recent theoretical and empirical analyses have demonstrated that, under the DPPS framework, one can obtain MAP-style or approximate posterior samples with improved measurement consistency, reduced sample variance, and strong theoretical guarantees for both linear and nonlinear observation models, including in cases where traditional likelihood score approximations fail or diversity is limited (Xu et al., 31 Jan 2025, Xu et al., 2024, Wu et al., 2024).

1. Theoretical Foundations

DPPS originates from the goal of sampling from the posterior

$p(x_0\mid y) \propto p(y|x_0)\, p(x_0)$

where $p(x_0)$ is the data prior modeled via a pretrained diffusion (score-based) generative model, and $p(y|x_0)$ encodes the measurement process, which may be linear, nonlinear, or even non-Gaussian (Xu et al., 31 Jan 2025, Xu et al., 2024, Li et al., 2023). The negative log-posterior decomposes as

$-\log p(x_0\mid y) = \ell(x_0) + r(x_0)$

with $\ell(x_0) = -\log p(y|x_0)$ the data-fidelity term and $r(x_0) = -\log p(x_0)$ the prior regularization.

At the core of DPPS is the exploitation of proximal operator theory: the posterior gradient can be written as the sum of likelihood- and prior-derived terms,

$$\nabla_{x_0} \log p(x_0|y) \approx -J_f(x_0)^\top (f(x_0)-y)/\sigma_y^2 + s_\theta(x_0,0)$$

where $s_\theta(x_0,0)$ is the diffusion model's score at $t\to 0$ (i.e., $\nabla_{x_0} \log p(x_0)$) and $J_f$ is the Jacobian of $f$ (Xu et al., 31 Jan 2025). In high dimensions and for Gaussian models, this enables computation of the MAP estimator or its stochastic generalization.

DPPS thus interprets the posterior sampling process as iterative proximal update dynamics, alternating between steps that improve data-consistency (via likelihood or measurement projection) and diffusion steps that move along the prior manifold (Wu et al., 2024, Xu et al., 31 Jan 2025, Montanari et al., 2023).

2. Algorithmic Structure and Variants

A typical DPPS iteration operates as follows (Xu et al., 31 Jan 2025, Wu et al., 2024, Xu et al., 2024):

1. Diffusion Reverse Step: Execute a standard DDPM or score-based denoising update to yield a sample consistent with the prior. For $x_t$ at timestep $t$, sample $x_{t-1}$ from the learned reverse kernel:

$$x_{t-1} \sim \mathcal{N}(\mu_{t-1}, \sigma_t^2 I), \quad \mu_{t-1} = \frac{1}{\sqrt{\alpha_t}}(x_t + \beta_t s_\theta(x_t, t))$$

2. Proximal (Measurement Consistency) Update: At each diffusion step or after several, apply a proximal MAP-like update:

$$z^{(j+1/2)} = z^{(j)} + \lambda \nabla_{x} \log p(y|z^{(j)})$$

$$z^{(j+1)} = \mu_{t-1} + r_t \frac{z^{(j+1/2)} - \mu_{t-1}}{\|z^{(j+1/2)} - \mu_{t-1}\|}$$

where the update is projected onto a sphere to maintain consistency with the conditional diffusion prior (Xu et al., 31 Jan 2025).

1. Candidate Selection: Multiple candidate samples per step (e.g., $K$ or $n$) are drawn; the candidate best matching data consistency (e.g., minimal $\|A \mu_0^{(j)} - y\|_2^2$) is chosen for the next iteration (Wu et al., 2024).
2. Lightweight Conditional Score Model (Optional): DPPS can warm-start via a small, control-conditioned score model trained on limited measurement/image pairs (e.g., using a ControlNet) (Xu et al., 31 Jan 2025).

Algorithmic pseudocode for canonical DPPS is provided in (Xu et al., 2024, Xu et al., 31 Jan 2025) and tailored for linear/nonlinear, single/multi-image, and high-dimensional inverse problems.

3. Connections to MAP Estimation, Posterior Sampling, and Related Methods

DPPS bridges deterministic MAP estimation and full posterior sampling through its hybrid use of score-based denoising and proximal consistency. Empirical and theoretical analysis shows that canonical DPS-type algorithms—based on an approximate conditional score—frequently produce samples with lower variance and diversity, thus behaving closer to MAP optimizers than true posterior samplers, particularly when the score is corrupted or miscalibrated (Xu et al., 31 Jan 2025). By explicitly formulating the measurement consistency step as a proximal operator (as in Moreau-Yosida regularization), DPPS achieves greater flexibility for both deterministic and stochastic sampling objectives, and it allows rigorous analysis and guarantees of convergence, bias, and sample diversity (Xu et al., 2024, Montanari et al., 2023, Ehrhardt et al., 2023).

This framework is distinct from projection-based solvers (e.g., ILVR, DDRM)—which enforce measurement constraints via hard projections—or from classical MCMC, which lack efficient high-dimensional priors. DPPS provides a modular plug-in approach, supporting a wide class of noisily observed or even quantized/mechanically degraded data models, and allowing for analytical computation of stepwise or asymptotic bias and variance (Xu et al., 2024, Montanari et al., 2023, Ehrhardt et al., 2023).

4. Generalizations and Extensions

DPPS generalizes naturally to several complex inverse settings:

• Multi-Image MRI Super-Resolution: In the MISR regime, the likelihood separability allows DPPS to sum independent measurement gradients and select proposals based on a joint, weighted data-fidelity cost. Substantial gains (1–2 dB PSNR, 0.01–0.02 SSIM) over single-image DPPS are observed, especially for anisotropic MRI (Remedios et al., 20 Jan 2026).
• Nonlinear Forward Models: For inverse problems such as nonlinear CT reconstruction with Poisson data, DPPS computes the posterior score via Bayes’ rule, fusing the pre-trained score model with an exact data-likelihood score, including the chain rule Jacobian contributed by a nonlinear forward operator (Li et al., 2023).
• High-Dimensional Statistics: DPPS has been formally connected to SDEs in parameter space (e.g., via AMP/TAP drift oracles), providing polynomial-time convergence guarantees and characterizing Wasserstein error in high dimensions (Montanari et al., 2023).
• Inexact Proximal Oracles: In scenarios where the proximal mapping cannot be computed exactly (e.g., with non-smooth or high-dimensional priors), the impact of errors on convergence and bias is quantitatively bounded, and practical stopping criteria are established (Ehrhardt et al., 2023).

5. Empirical Results and Performance Comparisons

DPPS delivers significant improvements over classical DPS, random-sample diffusion solvers, and even supervised conditional diffusion methods in various benchmarks:

Method	PSNR (↑)	SSIM (↑)	LPIPS (↓)	FID (↓)	Notes
DPS (orig.)	22.33 (SR×8)	–	0.4137	58.48	(Xu et al., 31 Jan 2025)
DPPS (K=2)	23.37	–	0.3770	44.37	Same runtime
DPPS (K=3, full)	23.52	–	0.3494	39.56	Higher runtime
DPPS (img rest.)	26.93 (SR×4, FFHQ)	0.778	0.203	26.30	(Wu et al., 2024)

DPPS achieves lower FID and LPIPS, crisper details, and fewer artifacts at negligible extra computation. Sample diversity is reduced compared to fully trained conditional diffusion models, consistent with the MAP-style update structure (Xu et al., 31 Jan 2025, Wu et al., 2024). In nonlinear CT, DPPS matches or slightly outperforms analytic and linearized DPS in PSNR and SSIM, and uniquely adapts to arbitrary measurement designs (Li et al., 2023).

6. Theoretical Guarantees and Practical Considerations

Rigorous analysis confirms that DPPS converges to the desired posterior under mild assumptions on model smoothness, convexity, and step size selection. Asymptotic consistency and non-asymptotic robustness theorem bounds guarantee that, as iteration count increases and annealing schedule decays, DPPS samples approach the true posterior distribution in total variation and Wasserstein distance (Xu et al., 2024, Ehrhardt et al., 2023).

Implementation requires careful choice of diffusion schedule, proximal step size, number of candidates in candidate selection, and optional conditioning strategies. Approximate or inexact solutions, including those due to finite MCMC steps or inner optimization tolerances, have bounded and quantifiable effects on output distributions (Ehrhardt et al., 2023). Lightweight conditional score models (e.g., ControlNet augmentations) can further improve sample quality with minimal extra training (Xu et al., 31 Jan 2025).

7. Broader Impact and Future Directions

DPPS provides a modular, robust, and theoretically sound backbone for posterior inference with expressive diffusion priors. Recent research highlights its adaptability to multi-measurement, nonlinear, and high-dimensional imaging problems, with extension potential to further contexts such as latent-space inference, distributed tomography, and medical reconstruction tasks.

Ongoing directions include tighter theoretical bounds (especially for non-convex and highly nonlinear problems), acceleration schemes (ordered-subsets, fast diffusion solvers), integration with learned conditional score estimation at scale, and exploration of diversification strategies to mitigate the inherent reduction in output diversity observed in MAP-centric update schemes (Xu et al., 31 Jan 2025, Li et al., 2023).

---

References:

• "Rethinking Diffusion Posterior Sampling: From Conditional Score Estimator to Maximizing a Posterior" (Xu et al., 31 Jan 2025)
• "Diffusion Posterior Proximal Sampling for Image Restoration" (Wu et al., 2024)
• "Provably Robust Score-Based Diffusion Posterior Sampling for Plug-and-Play Image Reconstruction" (Xu et al., 2024)
• "Likelihood-Separable Diffusion Inference for Multi-Image MRI Super-Resolution" (Remedios et al., 20 Jan 2026)
• "Posterior Sampling in High Dimension via Diffusion Processes" (Montanari et al., 2023)
• "Proximal Langevin Sampling With Inexact Proximal Mapping" (Ehrhardt et al., 2023)
• "CT Reconstruction using Diffusion Posterior Sampling conditioned on a Nonlinear Measurement Model" (Li et al., 2023)