Program

The conventional view in computational imaging casts noise as a nuisance to be minimized; however, recent advances suggest a different role for it entirely. Rather than something to suppress, deliberately injected stochasticity has driven reconstruction, regularization, and uncertainty quantification in ill-posed inverse problems. Score-based and diffusion priors, stochastic plug-and-play methods, and Langevin-type sampling all exploit controlled noise to explore posterior landscapes that deterministic methods struggle to capture. Tweedie's formula ties these together, linking Bayesian denoising and generative score-based and clarifying why a well-chosen noise process can be a computational asset rather than a liability. This special session brings together recent work that treats noise as a resource, spanning algorithmic frameworks, theory and convergence, posterior sampling and uncertainty quantification, with applications across a range of imaging modalities. By drawing connections across these threads, the session aims to clarify when and why stochasticity helps, and to surface the open questions shaping the field's next few years.

The design of computational imaging systems traditionally demands deep, cross-disciplinary expertise spanning physics, optics, mathematics, and algorithm development. This results in an R&D process that is slow, costly, and difficult to transfer across modalities, systems, and problem domains. This special session invites the community to explore a transformative new paradigm: the use of AI agents, including large language models (LLMs), to automate, accelerate, and co-design computational imaging systems. This session presents a series of talks followed by an interactive “flash panel” formed by conference participants. Each speaker will describe how they used agentic coding to solve major and/or minor problems in computational imaging, ranging from quick demos (vibes) to the development of accurate, robust methods and code (verification). The flash panel will then brainstorm for open problems and effective solutions in the age of agentic coding.

This session will focus on posterior sampling methods for computational imaging, with an emphasis on how generative models can be used to infer plausible images and quantify uncertainty in highly underdetermined inverse problems. The session will bring together perspectives on theory, algorithms, and applications, including diffusion and related generative approaches, uncertainty-aware reconstruction, and the role of learned priors in imaging. Together, the talks will highlight both current methodological advances and emerging opportunities and challenges for posterior sampling in computational imaging.

Event cameras represent a fundamentally different imaging paradigm in which individual pixels asynchronously report changes in log intensity rather than recording conventional frames. Their high temporal resolution, wide dynamic range, low latency, and sparse output enable imaging under rapid motion, extreme illumination changes, and bandwidth constraints. This special session will explore how event-based and neuromorphic sensing can push the limits of computational imaging, emphasizing connections among sensor physics, measurement models, hardware characterization, reconstruction algorithms, and system design. Topics may include event-generation and noise models; calibration and sensor nonidealities; information-theoretic analysis; reconstruction of intensity, motion, depth, and scene geometry; sparse and continuous-time representations; hybrid event-and-frame sensing; neuromorphic processing; joint sensor–algorithm optimization; and applications that exploit the unique capabilities of event sensors. By bringing together researchers in optics, computational imaging, hardware, control, computer vision, and machine learning, the session aims to foster cross-disciplinary discussion and identify opportunities for new forward models, reconstruction techniques, sensing architectures, and multimodal imaging systems.

Imaging problems in medicine and the sciences are rarely posed over a single image any more. A diagnosis draws on prior studies, clinical notes, and structured metadata; a materials characterization combines several instruments and length scales; a developmental study follows the same subject across years. The measurement is one channel among several, and the interesting modeling questions increasingly concern how those channels are combined. This session brings together work on multi-modal and vision–language models for medical and scientific imaging, organized around three recurring difficulties. The first is fusion: how to integrate sources of genuinely different character — images, text, structured variables, longitudinal observations — without collapsing them into an undifferentiated representation, and how to stay stable when a source is noisy, misaligned, or missing outright. The second is structure and interpretability: representations that separate what is shared across modalities from what is specific to each, and what that separation buys when a result has to be explained to a clinician or a domain scientist. The third is efficiency and deployment: compact models that run under real computational and instrument constraints, where accuracy matters only within a budget. The application domains are deliberately varied — radiology, ophthalmology, neuroimaging, microscopy, and large-scale scientific and synchrotron imaging — because the underlying formulations recur across all of them. The session aims to make those shared formulations visible across communities that do not usually meet, and to ask which ideas transfer and which need rethinking.

This session focuses on modern challenges in inverse problems involving coherent X-ray light, typically from synchrotron source but not exclusively; near and far field ptychography, phase contrast imaging, holography and their extension to three dimensions through tomography. We will focus on bridging the algorithmic design and challenges unique to practical applications arising in high resolution imaging, such as phase retrieval, large data, partially coherent illumination, structured illumination, geometrical errors and other stability issues. The workshop focuses on scalable mathematical approaches and algorithmic design; in particular pros and cons between different algorithms such as Conjugate Gradient, L-BFGS, ADMM, Alternating optimization etc in realistic scenarios in which the optimization functional contain many variables of different nature (object, probe, position errors etc) as well as preconditioning techniques.

The growing inflation in scientific publication enabled by advances in AI reflects a misdirection of this technology. This special session addresses the real uses of Machine Learning, Optimization, and Data Science to advance Materials Science and Microscopy through the common discipline of Computational Imaging, in a true peer-to-peer environment dedicated to advancing materials science and microscopy through the infusion of cross-disciplinary technologies. Relevant topics include computational imaging and inverse problems; optimization methods for imaging, reconstruction, and inference; AI- and ML-augmented microscopy; physics-informed machine learning; microstructure characterization, representation, and high-dimensional latent spaces; surrogate and learned physical models; identifiability, information content, ambiguity, and non-uniqueness; rare events and uncertainty quantification; large, heterogeneous, and multimodal materials data; fusion of imaging modalities, length scales, simulations, and physical measurements; imaging and prediction of evolving microstructures; non-ideal material response; contrast mechanisms; experimental design and autonomous imaging; generative and diffusion models; agentic AI and closed-loop workflows; structural artifacts and unsupported reconstructions; human interpretation of computationally augmented images; efficient computational methods; edge cases and instructive failures that reveal the limits of our methods; and new sources of complex imaging data enabled by autonomous measurements.

Neutron imaging instruments are tightly coupled systems built from highly specialized hardware and software. Because neutron sources are unique and complete off-the-shelf solutions rarely exist, it is difficult to develop instrument capabilities by designing the hardware first and adding the computation later. Detector physics, system control, data acquisition, instrument calibration, data reduction and reconstruction, and scientific interpretation have to be developed together across the instrument life-cycle. This session examines the neutron imaging instrument as a co-designed computational system and asks how interdependent hardware and software choices determine measurement performance, quantitative accuracy, extensibility, and scientific validity. Representative case studies will show how physics-based simulation can guide instrument optics, commissioning, and upgrades; how experiment-control and steering software can be developed alongside the hardware it operates; how commercial detector platforms can be adapted in both hardware and software to meet facility-specific requirements; and how hardware-aware reduction and reconstruction algorithms can exploit measurement physics to recover quantitative sample information and assess the corresponding measurement uncertainty. The speakers will address practical questions. Which aspects of an instrument require simulation during construction and upgrade? Which physical characteristics must be faithfully represented in software to produce physically accurate results? How should hardware and software changes be coordinated and validated to ensure scientifically correct interpretation? By placing neutron imaging in conversation with the broader computational-imaging community, the session aims to identify co-design principles that apply wherever sources are unique, instruments are custom-built, data are high-dimensional, and algorithms are critical to the measurement itself.