prNet: Data-Driven Phase Retrieval via Stochastic Refinement

arXiv:2507.09608 · eess.IV, cs.CV · Submitted 2025-07-13 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Electrical Engineering. The day in electrical engineering: signals, systems and the hardware that carries them.

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "prNet: Data-Driven Phase Retrieval via Stochastic Refinement".

Dev: Here's the gist of the paper. Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

Rosa: We propose a novel framework for phase retrieval that leverages Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality.

Dev: Unlike conventional approaches that prioritize pixel-wise accuracy, our methods navigate the perception-distortion tradeoff through a principled combination of stochastic sampling, learned denoising, and model-based updates.

Rosa: The Problem Context Phase retrieval involves reconstructing an unknown signal from noisy intensity measurements where the problem is highly non-linear and illposed.

Paper summary: Rosa: So to wrap up, the paper prNet: Data-Driven Phase Retrieval via Stochastic Refinement proposes a new way to handle phase retrieval by using Langevin dynamics for sampling from the posterior distribution.

Dev: The authors claim this framework yields reconstructions that explicitly balance distortion and perceptual quality, moving beyond methods that just try to be pixel-wise accurate.

Taro: It seems like the core contribution is structuring the problem into three variants, prNet-Small, prNet-Large, and prNet-Large-Adv.

Rosa: PrNet-Small uses Langevin dynamics with learned noise and denoising processes along with model driven HIO updates as its foundation.

Dev: PrNet-Large uses parallel sampling of diverse reconstructions to approximate the mean squared error solution through ensemble averaging to improve distortion metrics.

Taro: And prNet-Large-Adv builds on that by adding an adversarially trained aggregation network to specifically enhance perceptual quality while trying to maintain fidelity.

Rosa: The big picture is that this approach uses a principled combination of stochastic sampling and learned denoiser models within the Langevin dynamics framework to tackle the ill-posed nature of phase retrieval.

Dev: It suggests that combining these model-based updates with data-driven priors in a stochastic refinement framework could lead to more reliable solvers for nonlinear inverse problems.

Taro: For someone just listening, this means there's a new direction for how we tackle these hard inverse problems, focusing on the trade off between what the math tells you and what looks realistic.

Conclusion: Rosa: So, to wrap up this discussion about prNet: Data-Driven Phase Retrieval via Stochastic Refinement, we're looking at what these authors actually achieved here and what that means for us out here in the real world.

Dev: The title itself is a mouthful, but essentially, they’ve put together a framework that uses Langevin dynamics to sample from the probability distribution of the signal.

Taro: That’s right, it’s about using these stochastic updates to find a better reconstruction than what you get from just one shot of data.

Rosa: They tackle that problem by building three versions: prNet-Small, prNet-Large, and then the prNet-Large-Adv version which adds a little extra sharpening.

Dev: The core idea is balancing how accurate the image looks—the fidelity—with how realistic it actually seems to an observer.

Taro: It’s about navigating that distortion versus perceptual quality tradeoff, not just chasing pixel perfect noise reduction.

Rosa: They show that this combination of sampling and learned denoisers works really well across different benchmarks for phase retrieval.

Dev: The results are consistent: they beat the other methods on both traditional error metrics and those measures of how realistic the image looks to a human eye, like the SSIM score.

Taro: What’s interesting is that it doesn't introduce those weird artifacts you sometimes see when you push these models too far.

Rosa: They kept the runtime pretty reasonable compared to other complex setups, which is important for practical application on real hardware.

Dev: So, the authors suggest that coupling model-based updates with these stochastic refinement methods is a promising path for solving hard nonlinear problems in imaging.

Taro: It opens up a way for us to build more robust solvers when the world throws some weird data at us.

Rosa: And it sets up a question: can we take this framework and test it outside of the clean lab environment?

Middle East Technical University · Technical University of Denmark

eess.IV, cs.CV

Submitted: 2025-07-13

Updated: 2026-11-05

Comments: Accepted at IEEE Transactions on Computational Imaging (2026)

Journal ref: IEEE Transactions on Computational Imaging, vol. 12, pp. 1376-1391, 2026

DOI: 10.1109/TCI.2026.3719489

Code: https://github.com/METU-SPACE-Lab/prNet-forPhase-Retrieval

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 92/100

The gist: Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

Key concepts

Langevin Dynamics
This is a stochastic sampling method used to explore the complex probability distribution of possible solutions for phase retrieval. By incorporating learned noise and a denoiser model, it allows the algorithm to efficiently sample from the posterior distribution p(x|y), leading to more robust reconstructions than traditional methods.
Perception-Distortion Tradeoff
This is the central challenge in phase retrieval: finding a solution that is both accurate (high fidelity) and visually realistic (perceptual quality). The framework addresses this by strategically combining stochastic sampling, learned denoising, and model updates to find a principled balance between these two competing goals.
Hybrid Input-Output (HIO) Algorithm
The HIO algorithm is used within the Langevin dynamics update step to improve convergence properties. It allows the process to start from a good initial guess derived from classical solvers, speeding up the overall reconstruction process and enhancing stability.

Terminology

Summary

Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism. p. 0

Where the paper says it

1 prNet: Data-Driven Phase Retrieval via Stochastic Refinement Mehmet Onurcan Kaya and Figen S. Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

We propose a novel framework for phase retrieval that leverages Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality. p. 0
Where the paper says it

Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism. We propose a novel framework for phase retrieval that leverages Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality.

Unlike conventional approaches that prioritize pixel-wise accuracy, our methods navigate the perception-distortion tradeoff through a principled combination of stochastic sampling, learned denoising, and model-based updates. p. 0
Where the paper says it

We propose a novel framework for phase retrieval that leverages Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality. Unlike conventional approaches that prioritize pixel-wise accuracy, our methods navigate the perceptiondistortion tradeoff through a principled combination of stochastic sampling, learned denoising, and model-based updates.

The Problem Context

Phase retrieval involves reconstructing an unknown signal from noisy intensity measurements where the problem is highly non-linear and illposed. Classical approaches like the Error Reduction (ER) algorithm are sensitive to initialization and often stagnate at suboptimal solutions. p. 0

Where the paper says it

Among the earliest of these is the Error Reduction (ER) algorithm [3], which strictly applies hard projections at each step. While simple and computationally efficient, ER is highly sensitive to initialization and often stagnates at suboptimal solutions.

Deep learning methods often suffer from domain shift and tend to produce overly smooth outputs with reduced perceptual fidelity due to the perception-distortion tradeoff. p. 0
Where the paper says it

Such estimators tend to produce overly smooth outputs with reduced perceptual fidelity due to the perception-distortion tradeoff [14]. To address these limitations, deep generative models, particularly score/diffusion-based approaches, have gained traction for their ability to sample from the posterior distribution for many inverse problems [15].

Framework Components

The proposed framework consists of three progressively more advanced methods: (1) prNet-Small, which integrates Langevin dynamics with learned noise and denoising processes and model-driven HIO updates. (2) prNet-Large, which exploits parallel sampling of diverse reconstructions to approximate the MMSE solution via ensemble averaging. p. 1

Where the paper says it

We propose a novel diffusion-based framework for phase retrieval that bridges model-driven and data-driven approaches. Our framework consists of three progressively more advanced methods: (1) prNet-Small, a theoretically grounded and lightweight pipeline that integrates Langevin dynamics with learned noise and denoising processes and model-driven HIO updates; (2) prNet-Large, which exploits parallel sampling of diverse reconstructions to approximate the MMSE solution via ensemble averaging, significantly improving distortion metrics; and (3) prNet-Large-Adv, extending prNet-Large with an adversarially-trained aggregation network to enhance perceptual quality while preserving fidelity.

(3) prNet-Large-Adv, which extends prNet-Large with an adversarially trained aggregation network to enhance perceptual quality while preserving fidelity. p. 1
Where the paper says it

We propose a novel diffusion-based framework for phase retrieval that bridges model-driven and data-driven approaches. Our framework consists of three progressively more advanced methods: (1) prNet-Small, a theoretically grounded and lightweight pipeline that integrates Langevin dynamics with learned noise and denoising processes and model-driven HIO updates; (2) prNet-Large, which exploits parallel sampling of diverse reconstructions to approximate the MMSE solution via ensemble averaging, significantly improving distortion metrics; and (3) prNet-Large-Adv, extending prNet-Large with an adversarially-trained aggregation network to enhance perceptual quality while preserving fidelity.

Stochastic Refinement Mechanism

The core of the method is based on Langevin dynamics to sample from the posterior distribution p(xy) using both the score function and injected Gaussian noise. This involves an update step derived from Bayes’ rule for the posterior score function, which incorporates a learned denoiser model Dθ to approximate the unconditional score function. p. 4

Where the paper says it

Hence our approach is based on the following update step [15]: xt+1 ← xt + γ∇x log p (xt|y) + p 2γvt, 1 ≤ t ≤ T, (6) where γ denotes the step size, vt ∼ N (0, I) is standard Gaussian noise, and T is the total number of iterations. Using Bayes’ rule for the score function of the posterior yields: xt+1 ← xt+γ∇x log p (xt)+γ∇x log p (y|xt)+p 2γvt (7) We use a learned denoiser model Dθ with parameters θ to approximate the unconditional score function, assuming degradation with xt ≜ x + v where v ∼ N (0, σ2 t I) [47]: ∇x log p (xt) = Dθ (xt, t) − xt σ 2 t .

The update equation is then modified to incorporate the Hybrid Input-Output (HIO) algorithm for better convergence properties and warm-start initialization from classical solvers. p. 5
Where the paper says it

6 To address this, we can substitute this step with the Hybrid Input-Output (HIO) algorithm, which demonstrates better convergence properties in practice. This important improvement [12], [30] is often overlooked in prior diffusion-based PR methods such as [18], which typically follow simpler update rules with limited convergence behavior.

Performance Enhancements

The prNet-Large pipeline enhances reconstruction quality through multiple parallel reconstructions, where each estimate is passed through a denoiser to yield refined outputs. The prNet-Large-Adversarial pipeline further incorporates a second denoiser with a Wasserstein loss to enhance perceptual quality alongside distortion metrics. p. 7

Where the paper says it

To train the final denoiser Dϕ in the prNet-LargeAdversarial pipeline, we incorporate an additional improved Wasserstein GAN loss with gradient penalty [65] term into the training objective. This term addresses perceptual quality alongside distortion metrics, helping to balance the perceptiondistortion tradeoff.

Additionally, Test Time Augmentation (TTA D4) is employed to maximize the use of symmetry properties and available data by averaging results from all transformations in the D4 dihedral group. p. 8
Where the paper says it

By incorporating all transformations from the D4 dihedral group, this advanced TTA technique maximizes the use of symmetry properties and available data, significantly enhancing the robustness and quality of image reconstructions. This approach is particularly effective in image reconstruction tasks, where the enriched data from augmentation helps mitigate overfitting and improves generalization performance.

Experimental Validation

Extensive experiments demonstrate that our methods achieve state-of-the-art performance across multiple benchmarks, both in terms of fidelity and perceptual quality. The developed approach consistently surpasses other methods in both PSNR and SSIM metrics across all noise levels, while only necessitating a marginal increase in runtime compared to the initialization stage. p. 8

Where the paper says it

Table I presents the average reconstruction performance of the algorithms for 236 test images over 5 Monte Carlo runs under varying levels of Poisson noise (α = 2, 3, 4). The developed methods consistently surpass other methods in both PSNR and SSIM metrics across all noise levels, while only necessitating a marginal increase in runtime compared to the initialization stage.

Our methods generally do not introduce artifacts and errors like the other methods. p. 8
Where the paper says it

With our methods, HIO artifacts can be successfully removed while preserving the image characteristics. Our approach generally does not introduce artifacts and errors like the other methods.

Conclusion

Our results suggest that the combination of denoisers with model-based methods in the Langevin dynamics framework shows promise for developing reliable stochastic solvers for nonlinear inverse problems. The paper introduces a novel approach to phase retrieval using Langevin dynamics for posterior sampling. p. 0

Where the paper says it

Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism. We propose a novel framework for phase retrieval that leverages Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality.

The proposed methods consistently outperform both classical and modern baselines, while maintaining low computational runtime. p. 8
Where the paper says it

Table I presents the average reconstruction performance of the algorithms for 236 test images over 5 Monte Carlo runs under varying levels of Poisson noise (α = 2, 3, 4). The developed methods consistently surpass other methods in both PSNR and SSIM metrics across all noise levels, while only necessitating a marginal increase in runtime compared to the initialization stage.

The results suggest that the combination of denoisers with modelbased methods in the Langevin dynamics framework shows promise for developing reliable stochastic solvers for nonlinear inverse problems. p. 1
Where the paper says it

Moreover, our framework demonstrates promise for developing reliable stochastic nonlinear inverse problem solvers, which could have broader implications beyond PR. Some preliminary results of this research were presented in [19].

Acknowledgments

This study was funded in part by Scientific and Technological Research Council of Turkey (TUBITAK) under the Grant Number 120E505.

References

[1] J. Dong, L. Valzania, A. Maillard, T.-a. Pham, S. Gigan, and M. Unser, “Phase retrieval: From computational imaging to machine learning,” IEEE Signal Processing Magazine, vol. p. 0

Where the paper says it

Index Terms—Phase retrieval, diffusion models, deep learning, inverse problems, computational imaging. INTRODUCTION P HASE retrieval (PR) is a fundamental inverse problem in many scientific and engineering disciplines, where the goal is to reconstruct a signal using only intensity measurements such as Fourier intensities.

40, no. 1, pp. 45–57, 2023 ref:2507.096081[1]

[2] C. Metzler, P. Schniter, A. Veeraraghavan, and R. Baraniuk, “PrDeep: Robust phase retrieval with a flexible deep network,” in ICML, 2018 ref:2507.096084[2]

[3] J. R. Fienup, “Reconstruction of an object from the modulus of its fourier transform,” Optics Letters, vol. p. 8

Where the paper says it

Our experiments focus on the classical Fourier phase retrieval problem, which involves recovering an image from the magnitude of its Fourier transform. We assess generalization capability and computational efficiency, and compare reconstruction quality against both classical and state-of-the-art phase retrieval algorithms.

3, no. 1, pp. 27–29, 1978 ref:2507.096084[3]

[4] J. R. Fienup, “Phase retrieval algorithms: A comparison,” Applied Optics, vol. p. 8

Where the paper says it

For comparison, we present the results for prDeep [2], HIO [3], DIR [29], and MBwDDP [12], [30]. Table I presents the average reconstruction performance of the algorithms for 236 test images over 5 Monte Carlo runs under varying levels of Poisson noise (α = 2, 3, 4).

21, no. 15, pp. 2758–2769, 1982 ref:2507.096084[4]

[5] S. Marchesini, “Invited article: A unified evaluation of iterative projection algorithms for phase retrieval,” Review of scientific instruments, vol. p. 1

Where the paper says it

Iterative Projection Techniques for Phase Retrieval Iterative projection techniques have become fundamental tools for phase retrieval. One of the earliest and most wellknown algorithms is the classical Gerchberg-Saxton (GS) algorithm [20], which iteratively applies magnitude constraints in both the spatial and measurement domains to reconstruct an unknown signal.

78, no. 1, 2007 ref:2507.096084[5]

[6] I. Waldspurger, A. d’Aspremont, and S. Mallat, “Phase recovery, maxcut and complex semidefinite programming,” Mathematical Programming, vol. p. 0

Where the paper says it

Nevertheless, both ER and HIO remain vulnerable to noise, artifacts, and local minima, particularly in high-dimensional or low-SNR regimes [5]. To mitigate these issues, more advanced techniques have been developed, including methods based on semidefinite programming and sparse regularization.

149, pp. 47–81, 2015 ref:2507.096084[6]

[7] K. Jaganathan, S. Oymak, and B. Hassibi, “Sparse phase retrieval: Convex algorithms and limitations,” in ISIT, 2013 ref:2507.096084[7]

[8] S. Lopez-Tapia, R. Molina, and A. Katsaggelos, “Deep learning ap- proaches to inverse problems in imaging: Past, present and future,” Digital Signal Processing, vol. p. 0

Where the paper says it

Index Terms—Phase retrieval, diffusion models, deep learning, inverse problems, computational imaging. INTRODUCTION P HASE retrieval (PR) is a fundamental inverse problem in many scientific and engineering disciplines, where the goal is to reconstruct a signal using only intensity measurements such as Fourier intensities.

119, p. 103 285, 2021 ref:2507.096084[8]

[9] K. Wang et al., “On the use of deep learning for phase recovery,” Light: Science & Applications, vol. p. 2

Where the paper says it

This hybrid approach, which combines neural networks with analytical methods, has demonstrably succeeded in various real-valued 2D reconstruction problems, including deconvolution, super-resolution, tomography, and phase retrieval [10], [13], [25]. A key advantage of learning-based direct inversion methods lies in their low computational complexity due to their feed-forward (non-iterative) nature, making them suitable for real-time imaging applications.

13, no. 1, 2024 ref:2507.096084[9]

[10] K. H. Jin, M. T. McCann, E. Froustey, and M. A. Unser, “Deep convolutional neural network for inverse problems in imaging,” IEEE Transactions on Image Processing, vol. p. 0

Where the paper says it

In recent years, deep learning has emerged as a powerful tool for solving various inverse problems in imaging, including phase retrieval [8], [9]. Data-driven approaches based on deep neural networks (DNNs) have demonstrated remarkable success in directly reconstructing images from measurements or refining initial estimates from classical methods [10].

26, pp. 4509–4522, 2016 ref:2507.096084[10]

[11] Y. Romano, M. Elad, and P. Milanfar, “The little engine that could: Regularization by denoising (red),” SIAM Journal on Imaging Sciences, vol. p. 2

Where the paper says it

A key advantage of learning-based direct inversion methods lies in their low computational complexity due to their feed-forward (non-iterative) nature, making them suitable for real-time imaging applications. In contrast to learning-based direct inversion, plug-andplay regularization, and unrolled learning methods embrace iterative strategies.

10, no. 4, pp. 1804–1844, 2017 ref:2507.096084[11]

[12] C¸. Is¸ıl and F. p. 2

Where the paper says it

As theoretically shown in [14] and empirically observed in [40], these estimates may deviate significantly from the natural image manifold, leading to reconstructions with overly smooth features. Interestingly, the work by Is¸ıl et al.

p. 2
Where the paper says it

As theoretically shown in [14] and empirically observed in [40], these estimates may deviate significantly from the natural image manifold, leading to reconstructions with overly smooth features. Interestingly, the work by Is¸ıl et al.

S. Oktem, “Deep plug-and-play hio approach for phase retrieval,” Applied Optics, vol. p. 0
Where the paper says it

Alternatively, model-based optimization schemes have been augmented with deep priors learned from data using the plug-and-play framework [11], [12]. However, existing deep learning solutions for phase retrieval often suffer from limited performance due to domain shift, which occurs when the training data and real-world test data follow different distributions, leading to degraded accuracy [13].

64, no. 5, A84–A94, 2025 ref:2507.096084[12]

[13] W. Zhang, Y. Wan, Z. Zhuang, and J. Sun, “What’s wrong with end-to-end learning for phase retrieval? p. 1

Where the paper says it

While achieving state-of-the-art performance for simpler inverse problems such as denoising [24], these methods face challenges with complex observation models, significant discrepancies between observations and the target image, or limited training data availability. Such end-to-end direct inversion schemes also exist for the phase retrieval

” In IS&T Electronic Imaging Symposium, 2024 ref:2507.096084[13]

[14] Y. Blau and T. Michaeli, “The perception-distortion tradeoff,” in CVPR, 2018 ref:2507.096084[14]

[15] B. Kawar, G. Vaksman, and M. Elad, “Snips: Solving noisy inverse problems stochastically,” in NeurIPS, 2021 ref:2507.096084[15]

[16] C. Luo, “Understanding diffusion models: A unified perspective,” ArXiv, 2022 ref:2507.096084[16]

[17] M. O. Kaya and F. p. 0

Where the paper says it

1 prNet: Data-Driven Phase Retrieval via Stochastic Refinement Mehmet Onurcan Kaya and Figen S. Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

p. 0
Where the paper says it

1 prNet: Data-Driven Phase Retrieval via Stochastic Refinement Mehmet Onurcan Kaya and Figen S. Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

S. Oktem, “Ddrm-pr: Fourier phase retrieval using denoising diffusion restoration models,” Applied Optics, vol. p. 0
Where the paper says it

Recent works have applied the diffusion framework to phase retrieval. For instance, DDRM-PR [17], leverages pretrained diffusion models to solve inverse problems, but its reliance on generic denoisers that do not learn the forward measurement model limits its adaptability to specific imaging setups.

64, no. 5, A95–A105, 2025 ref:2507.096084[17]

[18] S. Shoushtari, J. p. 2

Where the paper says it

[44] utilized deep conditional generative models, such as cGAN and conditional VAE, to solve the inverse problem of digital holography, showcasing the potential of data-driven approaches in handling optical aberrations. Shoushtari et al.

Liu, and U. Kamilov, “Diffusion models for phase retrieval in computational imaging,” in ACSSC, 2023 ref:2507.096084[18]

[19] M. O. Kaya and F. S. Oktem, “Prnet: Efficient and robust phase retrieval via stochastic refinement,” in 2025 IEEE 35th International Workshop on Machine Learning for Signal Processing (MLSP), 2025, pp. p. 0

Where the paper says it

1 prNet: Data-Driven Phase Retrieval via Stochastic Refinement Mehmet Onurcan Kaya and Figen S. Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

01–06 ref:2507.096084[19]

[20] R. W. Gerchberg and W. p. 1

Where the paper says it

Iterative Projection Techniques for Phase Retrieval Iterative projection techniques have become fundamental tools for phase retrieval. One of the earliest and most wellknown algorithms is the classical Gerchberg-Saxton (GS) algorithm [20], which iteratively applies magnitude constraints in both the spatial and measurement domains to reconstruct an unknown signal.

O. Saxton, “A practical algorithm for the determination of phase from image and diffraction plane pictures,” Optik, vol. p. 3
Where the paper says it

Moreover, the flexibility of diffusion models allows their adaptation to various types of data and measurement settings. Whether dealing with coded diffraction patterns, multi-plane intensity measurements, or different wavelengths, diffusion models can be trained to incorporate these variations, enabling a unified framework for phase retrieval across diverse applications [17], [18].

35, pp. 237–250, 1972 ref:2507.096084[20]

[21] Y. Shechtman, Y. C. Eldar, O. Cohen, H. N. Chapman, J. Miao, and M. Segev, “Phase retrieval with application to optical imaging: A contemporary overview,” IEEE Signal Processing Magazine, vol. p. 0

Where the paper says it

Index Terms—Phase retrieval, diffusion models, deep learning, inverse problems, computational imaging. INTRODUCTION P HASE retrieval (PR) is a fundamental inverse problem in many scientific and engineering disciplines, where the goal is to reconstruct a signal using only intensity measurements such as Fourier intensities.

32, no. 3, pp. 87–109, 2015 ref:2507.096084[21]

[22] J. Qian, C. Yang, A. Schirotzek, F. Maia, and S. Marchesini, “Efficient algorithms for ptychographic phase retrieval, in inverse problems and applications,” Contemp. p. 5

Where the paper says it

Given that classical phase retrieval algorithms can already produce a fairly accurate reconstruction, it is more efficient to leverage this intermediate solution to avoid wasting denoiser model capacity on early-stage reconstruction. Such “image-to-image” rather than “noise-to-image” paradigm has also been recently exploited for different inverse problems in imaging [59], [60], [61], [62].

Math, vol. 615, pp. 261–280, 2014 ref:2507.096084[22]

[23] A. Maiden, D. Johnson, and P. Li, “Further improvements to the ptychographical iterative engine,” Optica, vol. p. 2

Where the paper says it

[29] attributes this smoothing behavior to an unavoidable inherent limitation of DNNs in the context of phase retrieval. However, as long as reconstruction algorithms prioritize minimizing distortion metrics, such as mean squared error, we can only expect limited improvements in perceptual quality.

4, no. 7, pp. 736–745, 2017 ref:2507.096084[23]

[24] M. El Helou and S. Susstrunk, “Blind universal bayesian image denoising with gaussian noise level learning,” IEEE Transactions on Image Processing, vol. p. 5

Where the paper says it

2, a well-established framework renowned for its efficacy in image restoration tasks [64]. Notably, this implementation of UNet incorporates timestep information as an additional input that is intricately linked to the noise level of the input image.

29, pp. 4885–4897, 2020 ref:2507.096084[24]

[25] Y. Nishizaki, R. Horisaki, K. Kitaguchi, M. Saito, and J. Tanida, “Analysis of non-iterative phase retrieval based on machine learning,” Optical Review, vol. p. 2

Where the paper says it

This hybrid approach, which combines neural networks with analytical methods, has demonstrably succeeded in various real-valued 2D reconstruction problems, including deconvolution, super-resolution, tomography, and phase retrieval [10], [13], [25]. A key advantage of learning-based direct inversion methods lies in their low computational complexity due to their feed-forward (non-iterative) nature, making them suitable for real-time imaging applications.

27, pp. 136–141, 2020 ref:2507.096084[25]

[26] T. Uelwer, A. Oberstrass, and S. Harmeling, “Phase retrieval using conditional generative adversarial networks,” in ICPR, 2020 ref:2507.096084[26]

[27] I. Shevkunov, J. Kilpelainen, and K. Eguiazarian, “Deep convolutional ¨ neural network-based lensless quantitative phase retrieval,” in BiOS, 2021 ref:2507.096084[27]

[28] S. V. Venkatakrishnan, C. A. Bouman, and B. Wohlberg, “Plug-and-play priors for model based reconstruction,” in GlobalSIP, 2013 ref:2507.096084[28]

[29] C¸. Is¸ıl, F. p. 2

Where the paper says it

As theoretically shown in [14] and empirically observed in [40], these estimates may deviate significantly from the natural image manifold, leading to reconstructions with overly smooth features. Interestingly, the work by Is¸ıl et al.

S. Oktem, and A Koc¸, “Deep iterative reconstruction for phase retrieval,” Applied Optics, vol. p. 0
Where the paper says it

1 prNet: Data-Driven Phase Retrieval via Stochastic Refinement Mehmet Onurcan Kaya and Figen S. Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

58, no. 20, pp. 5422–5431, 2019 ref:2507.096084[29]

[30] C¸. Is¸ıl and F. S. Oktem, “Model-based phase retrieval with deep denoiser prior,” in COSI, 2020 ref:2507.096084[30]

[31] E. J. Cha, C. Lee, M. Jang, and J. C. Ye, “Deepphasecut: Deep relaxation in phase for unsupervised fourier phase retrieval,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. p. 0

Where the paper says it

1 prNet: Data-Driven Phase Retrieval via Stochastic Refinement Mehmet Onurcan Kaya and Figen S. Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism.

44, pp. 9931–9943, 2020 ref:2507.096084[31]

[32] Y. Wang, X. Sun, and J. W. Fleischer, “When deep denoising meets iterative phase retrieval,” in ICML, 2020 ref:2507.096084[32]

[33] H. K. Aggarwal, M. P. Mani, and M. Jacob, “Modl: Model-based deep learning architecture for inverse problems,” IEEE Transactions on Medical Imaging, vol. p. 0

Where the paper says it

In recent years, deep learning has emerged as a powerful tool for solving various inverse problems in imaging, including phase retrieval [8], [9]. Data-driven approaches based on deep neural networks (DNNs) have demonstrated remarkable success in directly reconstructing images from measurements or refining initial estimates from classical methods [10].

38, pp. 394–405, 2017 ref:2507.096084[33]

[34] V. Monga, Y. Li, and Y. C. Eldar, “Algorithm unrolling: Interpretable, efficient deep learning for signal and image processing,” IEEE Signal Processing Magazine, vol. p. 0

Where the paper says it

Oktem , Member, IEEE Abstract—Phase retrieval is an ill-posed inverse problem in which classical and deep learning–based methods struggle to jointly achieve measurement fidelity and perceptual realism. We propose a novel framework for phase retrieval that leverages Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality.

38, pp. 18–44, 2019 ref:2507.096084[34]

[35] N. Naimipour, S. Khobahi, and M. Soltanalian, “Upr: A model-driven architecture for deep phase retrieval,” ACSSC, 2020 ref:2507.096084[35]

[36] M. Deng, A. Goy, K. Arthur, and G. Barbastathis, “Physics embedded deep neural network for phase retrieval under low photon conditions,” in COSI, 2019 ref:2507.096084[36]

[37] N. Naimipour, S. Khobahi, M. Soltanalian, H. Safavi, and H. C. Shaw, “Unfolded algorithms for deep phase retrieval,” Algorithms, vol. p. 1

Where the paper says it

Iterative Projection Techniques for Phase Retrieval Iterative projection techniques have become fundamental tools for phase retrieval. One of the earliest and most wellknown algorithms is the classical Gerchberg-Saxton (GS) algorithm [20], which iteratively applies magnitude constraints in both the spatial and measurement domains to reconstruct an unknown signal.

17, no. 12, 2024 ref:2507.096084[37]

[38] C.-J. Wang, C.-K. Wen, S.-H. L. Tsai, and S. Jin, “Phase retrieval with learning unfolded expectation consistent signal recovery algorithm,” IEEE Signal Processing Letters, vol. p. 1

Where the paper says it

Iterative Projection Techniques for Phase Retrieval Iterative projection techniques have become fundamental tools for phase retrieval. One of the earliest and most wellknown algorithms is the classical Gerchberg-Saxton (GS) algorithm [20], which iteratively applies magnitude constraints in both the spatial and measurement domains to reconstruct an unknown signal.

27, pp. 780–784, 2020 ref:2507.096084[38]

[39] A. Liu, X. Fan, Y. Yang, and J. Zhang, “Prista-net: Deep iterative shrinkage thresholding network for coded diffraction patterns phase retrieval,” ArXiv, 2023 ref:2507.096084[39]

[40] C. Ledig et al., “Photo-realistic single image super-resolution using a generative adversarial network,” in CVPR, 2017 ref:2507.096084[40]

[41] K. Tayal, C.-H. Lai, R. Manekar, Z. Zhuang, V. Kumar, and J. Sun, “Unlocking inverse problems using deep learning: Breaking symmetries in phase retrieval,” in NeurIPS 2020 Workshop on Deep Learning and Inverse Problems, 2020 ref:2507.096084[41]

[42] A. G. Dimakis, “Deep generative models and inverse problems,” in Mathematical Aspects of Deep Learning, P. p. 2

Where the paper says it

Despite these limitations, unrolled learning has shown success in phase retrieval [35], [36], [37], [38], [39]. Generative Models for Inverse Problems All of the aforementioned deep learning methods focus on Maximum A Posteriori or Minimum Mean Squared Error (MMSE) estimation.

Grohs and G. Kutyniok, Eds. Cambridge University Press, 2022, pp. 400–421 ref:2507.096084[42]

[43] Z. Zhao, J. C. Ye, and Y. Bresler, “Generative models for inverse imaging problems: From mathematical foundations to physics-driven applications,” IEEE Signal Processing Magazine, vol. p. 0

Where the paper says it

Index Terms—Phase retrieval, diffusion models, deep learning, inverse problems, computational imaging. INTRODUCTION P HASE retrieval (PR) is a fundamental inverse problem in many scientific and engineering disciplines, where the goal is to reconstruct a signal using only intensity measurements such as Fourier intensities.

40, no. 1, pp. 148–163, 2023 ref:2507.096084[43]

[44] J. Gladrow, “Digital phase-only holography using deep conditional generative models,” ArXiv, 2019 ref:2507.096084[44]

[45] S. Chan, “Tutorial on diffusion models for imaging and vision,” Foundations and Trends in Computer Graphics and Vision, vol. p. 0

Where the paper says it

com/METU-SPACE-Lab/prNet-forPhase-Retrieval. Index Terms—Phase retrieval, diffusion models, deep learning, inverse problems, computational imaging.

16, no. 4, pp. 322–471, 2024 ref:2507.096084[45]

[46] X. Li et al., “Diffusion models for image restoration and enhancement, a comprehensive survey,” ArXiv, 2023 ref:2507.096084[46]

[47] K. Miyasawa, “An empirical bayes estimator of the mean of a normal population,” Bull. p. 3

Where the paper says it

While computationally efficient, requiring no modifications to the unconditional pretrained diffusion model’s score function, this approach inherits a fundamental limitation: the initialization does not enforce measurement consistency. Since sampling begins at an intermediate denoised state (skipping early noise-heavy diffusion steps) yet still relies on an unconditional score estimator, the generated samples may diverge from the true data manifold or violate physical constraints.

Inst. Internat. Statist, vol. 38, no. 181-188, pp. 1–2, 1961 ref:2507.096084[47]

[48] C. Meng et al., “SDEdit: Guided image synthesis and editing with stochastic differential equations,” in ICLR, 2022 ref:2507.096084[48]

[49] C. Saharia, J. Ho, W. Chan, T. Salimans, D. J.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Stochastic Refinement via Langevin Dynamics for Phase Retrieval: The system can perform phase retrieval by leveraging Langevin dynamics to enable efficient posterior sampling, yielding reconstructions that explicitly balance distortion and perceptual quality through a principled combination of stochastic sampling, learned denoising, and model-based updates.

  2. Progressive Framework for Model Complexity: The framework comprises three variants of increasing complexity: prNet-Small (theoretically grounded), prNet-Large (parallel reconstruction sampling to approximate the MMSE solution), and prNet-Large-Adv (adversarially trained aggregation network for enhanced perceptual quality).

  3. Integration of Hybrid Input-Output Updates: The system can incorporate the Hybrid Input-Output (HIO) algorithm into its updates, which demonstrates better convergence properties in practice and allows for the incorporation of object domain constraints such as real-valuedness and non-negativity.

  4. Test Time Augmentation for Robustness: The system can employ Test Time Augmentation (TTA), including the advanced TTA D4 technique, to enhance reconstruction quality by leveraging inherent invariances of measurement operators and maximizing symmetry properties through averaging over all D4 transformations.

  5. Learned Denoising Schedule: The system can learn both noise and denoising schedule during training, allowing the model to discover an optimal denoising strategy tailored to the reconstruction task, which improves performance compared to approaches relying on a fixed, predefined diffusion process.

  6. Adversarial Loss for Perceptual Realism: For prNet-Large-Adversarial, it can incorporate an improved Wasserstein GAN loss with gradient penalty term to enhance perceptual quality alongside distortion metrics, mitigating the smoothing that are prevalent in other methodologies.

Abstract

Phase retrieval is an ill-posed inverse problem in which existing methods often struggle to balance reconstruction fidelity and perceptual quality. We propose a novel Langevin-based framework for phase retrieval that bridges model-driven and data-driven approaches by combining stochastic posterior sampling, learned denoising, and model-based updates to better navigate the perception-distortion tradeoff. The framework consists of three progressively more advanced variants. The first, prNet-Small, provides a theoretically grounded and lightweight pipeline that integrates Langevin dynamics with learned noise and denoising processes and alternating projection-based updates. Building upon this foundation, prNet-Large exploits parallel sampling of diverse reconstructions and aggregates them to approximate the MMSE solution, significantly improving distortion metrics. Finally, prNet-Large-Adversarial enhances perceptual quality by replacing simple averaging with an adversarially trained aggregation network while preserving reconstruction fidelity. In addition, we incorporate test time augmentation to further improve the performance. Extensive experiments demonstrate that the proposed framework consistently outperforms classical, deep learning-based, and diffusion-based baselines across multiple benchmarks in both reconstruction fidelity and perceptual quality. The source code and trained models are available at https://github.com/METU-SPACE-Lab/prNet-for-Phase-Retrieval

Related papers