research

Reliable learning-based solvers for imaging inverse problems.

Modern imaging systems increasingly rely on learned priors (denoisers, diffusion models, and other generative models) to reconstruct images from incomplete or noisy measurements. These methods are powerful, but they are trained on one distribution and deployed on another. My research asks when such solvers can be trusted, how to detect when they cannot, and how to design new ones with guarantees.

Robustness under model mismatch

What happens to a learned reconstruction method when its prior or measurement model does not match the data it sees at test time? I study plug-and-play (PnP) and deep equilibrium architectures under mismatched priors and forward models. I characterize how mismatch affects convergence and reconstruction error, and develop lightweight adaptation strategies, such as domain adaptation and test-time training, that recover performance with little target data.

  1. ICML
    Prior Mismatch and Adaptation in PnP-ADMM with a Nonconvex Convergence Analysis
    Shirin Shoushtari*, Jiaming Liu*, Edward P. Chandler, and 2 more authors
    In International Conference on Machine Learning (ICML), 2024
    arXiv HTML Code Website
  2. arXiv
    PnP Restoration with Domain Adaptation for SANS
    Shirin Shoushtari, Edward P. Chandler, Jialiang Zhang, and 4 more authors
    arXiv preprint, 2024
    arXiv
  3. CAMSAP
    Overcoming Distribution Shifts in Plug-and-Play Methods with Test-Time Training
    Edward P. Chandler, Shirin Shoushtari, Jiaming Liu, and 2 more authors
    In IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2023
    arXiv HTML
  4. ICASSP
    Robustness of Deep Equilibrium Architectures to Changes in the Measurement Model
    Junhao Hu*, Shirin Shoushtari*, Zihao Zou, and 3 more authors
    In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2023
    arXiv HTML
  5. JSAIT
    Deep Model-based Architectures for Inverse Problems under Mismatched Priors
    Shirin Shoushtari, Jiaming Liu, Yuyang Hu, and 1 more author
    IEEE Journal on Selected Areas in Information Theory, 2022
    arXiv HTML Website

Detecting and measuring distribution shift

Before correcting for a shift, one has to know it is there. I develop methods that use diffusion models to detect out-of-distribution inputs and estimate distribution shift directly from measurements, without access to clean ground-truth images. These include covariance-based OOD scores and measurement-domain divergence estimates.

  1. ICLR
    EigenScore: OOD Detection using Covariance in Diffusion Models
    Shirin Shoushtari, Yi Wang, Xiao Shi, and 2 more authors
    In International Conference on Learning Representations (ICLR), 2026
    arXiv Code Website
  2. CAMSAP
    Distribution Shift Estimation in Imaging Inverse Problems
    Weining Wang*, Shirin Shoushtari*, Edward P. Chandler, and 2 more authors
    In IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2025
    HTML
  3. Asilomar
    Out-of-Distribution Detection via Uncertainty Quantification in Diffusion Models
    Shirin Shoushtari, Yi Wang, and Ulugbek S. Kamilov
    In Asilomar Conference on Signals, Systems, and Computers, 2025
  4. arXiv
    Unsupervised Detection of Distribution Shift in Inverse Problems using Diffusion Models
    Shirin Shoushtari, Edward P. Chandler, M. Salman Asif, and 1 more author
    arXiv preprint, 2025
    arXiv Code Website

Diffusion and generative priors for inverse problems

I work on using diffusion, flow-based, and other generative models as priors for imaging problems such as phase retrieval, Poisson denoising, and video restoration. This includes models that learn directly in the measurement domain, one-step posterior sampling for noisy measurements, and flow-based reconstruction through regularized minimization.

  1. arXiv
    SNaP: One-Step Posterior Sampling for Noisy Inverse Problems
    Shirin Shoushtari*, Edward P. Chandler*, Xiao Shi, and 1 more author
    arXiv preprint, 2026
    arXiv Website
  2. arXiv
    FIRM: Flow-based Imaging via Regularized Minimization
    Shirin Shoushtari, Edward P. Chandler, Xiao Shi, and 1 more author
    arXiv preprint, 2026
    arXiv Website
  3. ICLR
    Measurement Score-Based Diffusion Model
    Chicago Y. Park, Shirin Shoushtari, Hongyu An, and 1 more author
    In International Conference on Learning Representations (ICLR), 2026
    arXiv Code Website
  4. ICASSP
    Moments Matter: Posterior Recovery in Poisson Denoising via Log-Networks
    Shirin Shoushtari, Edward P. Chandler, and Ulugbek S. Kamilov
    In IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), 2026
    arXiv
  5. WACV
    FLAIR: A Conditional Diffusion Framework with Applications to Face Video Restoration
    Zihao Zou*, Jiaming Liu*, Shirin Shoushtari, and 3 more authors
    In IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025
    arXiv HTML Code Website
  6. Asilomar
    DOLPH: Diffusion Models for Phase Retrieval
    Shirin Shoushtari, Jiaming Liu, and Ulugbek S. Kamilov
    In Asilomar Conference on Signals, Systems, and Computers, 2023
    arXiv HTML

Optimization and convergence theory

Underlying all of this is optimization. I analyze the convergence of nonconvex PnP-ADMM with MMSE denoisers, develop closed-form proximal operators, and study analysis-based and block-coordinate PnP methods for blind and structured inverse problems.

  1. TCI
    Closed-Form Approximation of the Total Variation Proximal Operator
    Edward P. Chandler, Shirin Shoushtari, Brendt Wohlberg, and 1 more author
    IEEE Transactions on Computational Imaging, 2025
    arXiv HTML
  2. arXiv
    Analysis Plug-and-Play Methods for Imaging Inverse Problems
    Edward P. Chandler, Shirin Shoushtari, Brendt Wohlberg, and 1 more author
    arXiv preprint, 2025
    arXiv
  3. NeurIPS
    Block Coordinate Plug-and-Play Methods for Blind Inverse Problems
    Weijie Gan, Shirin Shoushtari, Yuyang Hu, and 3 more authors
    In Advances in Neural Information Processing Systems (NeurIPS), 2023
    arXiv HTML Code Website
  4. CAMSAP
    Convergence of Nonconvex PnP-ADMM with MMSE Denoisers
    Chicago Y. Park*, Shirin Shoushtari*, Weijie Gan, and 1 more author
    In IEEE International Workshop on Computational Advances in Multi-Sensor Adaptive Processing (CAMSAP), 2023
    Student Paper Award Finalist arXiv HTML Website
  5. NeurIPS
    Online Deep Equilibrium Learning for Regularization by Denoising
    Jiaming Liu, Xiaojian Xu, Weijie Gan, and 2 more authors
    In Advances in Neural Information Processing Systems (NeurIPS), 2022
    arXiv HTML Code

Looking ahead: flow matching with guarantees

Flow matching has emerged as a flexible alternative to diffusion models, and it has deep connections to optimization and minimization theory. My next goal is to exploit those connections to design principled flow-based solvers for imaging inverse problems with provable guarantees, combining the expressiveness of modern generative models with the reliability of classical optimization.