Generative Latent Diffusion Priors for Large-Scale Computational Imaging

by Dr Suman Saha

Wednesday 21 Jan 2026, 11:30 → 13:00 Europe/Zurich

OHSA/B17

Abstract:
Computational imaging problems in scientific domains are often ill-posed, data-limited, and constrained by expensive acquisition or iterative reconstruction pipelines. Classical optimization-based methods struggle to scale under these conditions, while purely discriminative models often lack robustness when data are scarce or incomplete. In this work, we present a generative modeling framework based on latent diffusion priors, designed to capture global structural regularities and serve as a reusable prior for large-scale computational imaging tasks. We demonstrate the effectiveness of this framework on two challenging applications at PSI. First, in 3D neuron segmentation for connectomics [1], we leverage generative affinity modeling to improve the 3D segmentation of complex neuronal morphologies in volumetric electron microscopy, enhancing robustness to ambiguous boundaries and long-range dependencies. Second, we apply the same generative principles to ptychographic phase retrieval for EUV photomask inspection, introducing Ptycho-LDM [2], a hybrid physics-informed pipeline where a conditional latent diffusion model refines coarse ptychographic reconstructions. Across both domains, our results show that generative latent diffusion priors provide a unifying and scalable approach for advancing computational imaging by effectively bridging data-driven learning and physics-based modeling.

References:
[1] Xiaoyu Liu, et al. Cross-Dimension Affinity Distillation for 3D EM Neuron Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.

[2] Suman Saha, Paolo Ansuinelli, Luis Barba, Iacopo Mochi, Benjamín Béjar Haro. Ptycho-LDM: A Hybrid Framework for Efficient Phase Retrieval of EUV Photomasks Using Conditional Latent Diffusion Models. Photonics, vol. 12, no. 9, 2025.

TEAMS LINK