PSI3D: Plug-and-Play 3D Stochastic Inference with Slice-wise Latent Diffusion Prior

2025-12-20

AI Summary

Traditional diffusion models struggle to scale to kilovoxel-scale 3D scientific imaging (e.g., 1024×1024×128 OCT volumes), suffering from prohibitive training and inference costs. Method: We propose a plug-and-play stochastic Bayesian inversion framework that couples slice-wise 2D latent diffusion sampling with inter-slice stochastic total variation (TV) regularization, enabling efficient and spatially consistent 3D reconstruction; theoretically, it guarantees Markov chain convergence and principled uncertainty quantification. Contribution/Results: This work is the first to overcome the scalability bottleneck of diffusion models for kilovoxel-scale 3D inverse problems. It enables statistically grounded, uncertainty-aware reconstruction. In OCT super-resolution, our method significantly outperforms conventional optimization- and learning-based baselines, achieving an unprecedented balance among reconstruction fidelity, computational feasibility, and statistical rigor.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Reasoning under Uncertainty: Stochastic Optimization
• Machine Learning: Calibration & Uncertainty Quantification

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• User Modeling, Personalization and Recommendation: On-Device user modeling, personalization, and recommendation
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

Diffusion models are highly expressive image priors for Bayesian inverse problems. However, most diffusion models cannot operate on large-scale, high-dimensional data due to high training and inference costs. In this work, we introduce a Plug-and-play algorithm for 3D stochastic inference with latent diffusion prior (PSI3D) to address massive ($1024 imes 1024 imes 128$) volumes. Specifically, we formulate a Markov chain Monte Carlo approach to reconstruct each two-dimensional (2D) slice by sampling from a 2D latent diffusion model. To enhance inter-slice consistency, we also incorporate total variation (TV) regularization stochastically along the concatenation axis. We evaluate our performance on optical coherence tomography (OCT) super-resolution. Our method significantly improves reconstruction quality for large-scale scientific imaging compared to traditional and learning-based baselines, while providing robust and credible reconstructions.

Problem

• Addresses large-scale 3D volume reconstruction using latent diffusion models
• Enhances slice consistency in 3D imaging via stochastic total variation regularization
• Improves reconstruction quality for scientific imaging like OCT super-resolution

Innovation

• Uses 2D latent diffusion model for slice-wise 3D reconstruction
• Incorporates stochastic total variation for inter-slice consistency
• Enables large-scale 3D imaging via plug-and-play Bayesian inference