Fourier phase retrieval aims to reconstruct images from magnitude-only measurements; however, the absence of phase information, noises, and channel ambiguities make the recovery of natural color images challenging. Although diffusion-prior-based methods impose strong image constraints, their recovery performance degrades under higher measurement noise, and repeated generative updates incur substantial computational overhead. To overcome these limitations, this paper proposes a novel method, termed FlowFPR, that integrates a pretrained flow-matching prior with the relaxed averaged alternating reflections (RAAR) algorithm for oversampled Fourier phase retrieval. In the basic version of FlowPR (Flow-RAAR-basic), a color image is estimated by repeatedly performing four steps at each outer iteration, where the flow-prior prediction step exploits single-step straight-line prediction, considerably reducing the generation cost. Nevertheless, Flow-RAAR-basic still suffers from two critical issues: (i) the predefined fusion weights, λn, might lead to either ineffective or excessive correction, and (ii) the fixed iteration number used by RAAR algorithm does not allow adaptation to measurement noise. To address these two issues, we develop two improved algorithms, Flow-RAAR-adaptive and Flow-RAAR-controlled, respectively. Moreover, these two improved algorithms naturally lead to a joint variant, Flow-RAAR-joint, which combines their advantages. Extensive experiments on a natural face image dataset demonstrate the superior performance of the four proposed methods compared with other relevant competing methods, as well as their acceleration over two recent diffusion-based methods. In the noise setting, Flow-RAAR-joint performs best among the four proposed methods.
Figures & tables
Fig. 1: FlowFPR reconstruction trajectory from the random initial state x0 to the estimate xN . Each outer iteration re-noises xn to obtain ztn , predicts x^n using the Flow prior, generates a RAAR candidate xnR , and fuses it with x^n to obtain xn+1 .
Parameter
Value
Image / observation size
256×256 / 512×512
Oversampling parameter κ
4
Outer iterations N
800
tmin,tmax
0.25,0.95
Baseline fusion schedule g(t)
t
λmin,λmax
0,1
TABLE I: Main Experimental Parameters of the FlowPR Methods
σ=0
σ=0.03
σ=0.05
Method
PSNR
SSIM
LPIPS
MSR (%)
PSNR
SSIM
LPIPS
MSR (%)
PSNR
SSIM
LPIPS
MSR (%)
HIO
21.3528
0.6409
0.4377
43.9
16.2323
0.2817
0.7540
2.3
14.7042
0.2011
0.8135
0.0
RAAR
18.5623
0.4349
0.6479
5.5
18.3807
0.3667
0.6976
4.9
18.2471
0.3232
0.7148
4.7
DPS
27.7673
0.7908
0.1333
42.4
24.9240
0.7104
0.1722
40.4
23.8213
0.6747
0.1898
39.9
DiffFPR
36.9518
0.9005
0.0779
68.4
25.3380
0.6032
0.3041
66.6
23.0191
0.4566
0.4589
65.0
Flow-RAAR-basic
47.1896
0.9932
0.0013
59.3
27.7261
0.6245
0.2447
55.2
24.3497
0.4558
0.4323
52.0
TABLE II: Mean Reconstruction Results on 50 FFHQ Images at Different Noise Levels
Fig. 2: Qualitative reconstructions under noiseless observations. Rows from top to bottom show images 19123, 27175, 03899, and 04561. Columns from left to right show ground truth, HIO, DPS, DiffFPR, Flow-RAAR-basic, Flow-RAAR-adaptive, Flow-RAAR-controlled, and the joint variant. DiffFPR follows the channelwise ambiguity-resolution rule in its public code.
Fig. 3: Qualitative reconstructions at σ=0.03 . Rows from top to bottom show images 38478, 22710, 38454, and 38844. The first two are randomly selected DiffFPR failures (PSNR below 20 dB), and the last two are successful cases (PSNR at least 20 dB). Method columns follow Fig. 2 .
Fig. 4: Qualitative reconstructions at σ=0.05 . Rows from top to bottom show images 04899, 45853, 00517, and 14560. The first two are randomly selected DiffFPR failures (PSNR below 20 dB), and the last two are successful cases (PSNR at least 20 dB). Method columns follow Fig. 2 .
Consistency operator
PSNR (dB)
SSIM
LPIPS
Relative residual
HIO
20.9394
0.4648
0.5608
0.15853
RAAR
27.6415
0.6279
0.2387
0.11369
TABLE III: Ablation of RAAR and HIO Data-Consistency Operators Within the Flow-Prior Framework ( σ=0.03 ).
σ
N
K
NK
PSNR
SSIM
LPIPS
Relative residual
0
400
10
4000
46.1400
0.9913
0.0012
0.00185
800
10
8000
46.9896
0.9934
0.0009
0.00175
1200
10
12000
46.9656
0.9930
0.0010
0.00176
800
5
4000
43.9924
0.9872
0.0023
0.00272
800
10
8000
46.9896
0.9934
0.0009
0.00175
800
15
12000
48.4409
0.9951
0.0006
0.00145
TABLE IV: Ablation of Outer Iterations N and Fixed Inner Steps K at Different Noise Levels.
σ
g(t)
PSNR (dB)
SSIM
LPIPS
Relative residual
0
t
47.0630
0.9930
0.00090
0.00158
t
46.9896
0.9934
0.00095
0.00175
t2
45.3022
0.9890
0.00490
0.00281
0.05
t
24.1720
0.4541
0.4357
0.19266
t
24.2539
0.4587
0.4303
0.19255
t2
24.3626
0.4661
0.4243
0.19233
TABLE V: Sensitivity to the Baseline Fusion Schedule g(t) at Different Noise Levels.
σ
ρ
PSNR (dB)
SSIM
LPIPS
Jˉ
jˉ⋆
0
0.85
42.2200
0.9762
0.0070
1.497
1.497
0.90
40.4453
0.9685
0.0119
1.210
1.210
0.95
38.5404
0.9604
0.0166
1.039
1.039
0.05
0.85
27.2900
0.6187
0.2156
8.155
6.673
0.90
27.6565
0.6354
0.1990
6.668
5.301
0.95
27.6370
0.6339
0.2026
3.822
2.750
TABLE VI: Sensitivity to the Flow-RAAR-controlled Threshold ρ at Different Noise Levels. Here, Jˉ is the mean number of inner iterations performed, and jˉ⋆ is the mean selected candidate index.
Stopping rule
PSNR
SSIM
LPIPS
Jˉ
jˉ⋆
Run to Kmax
28.3409
0.6758
0.1797
10.000
9.577
Early stopping
28.7110
0.6777
0.1731
5.047
4.654
TABLE VII: Ablation of the Flow-RAAR-controlled Early Stopping Rule ( σ=0.03 , Kmax=10 ).
Method
Total time (h)
Mean time (s/image)
Relative speedup
DiffFPR
5.229
753.04
1.00
DPS
8.953
1289.24
0.58
Flow-RAAR-basic
3.463
498.60
1.51
Flow-RAAR-adaptive
4.508
649.08
1.16
Flow-RAAR-controlled
3.410
491.08
1.53
Flow-RAAR-joint
3.812
548.91
1.37
TABLE VIII: Runtime Comparison on 25 Images Using the Same RTX 4070 SUPER, where DiffFPR is selected as the baseline.
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
Deep unrolling methods for inverse problems augment traditional reconstruction algorithms with learned priors to improve their performance. These methods have been developed under the implicit assumption that the geometry of the measurement operator does not change between training and deployment; when this is no longer the case, the quality of obtained reconstructions can degrade significantly. In this paper, we argue that this is not an inherent limitation of the deep unrolling paradigm and can be mitigated by a careful choice of the underlying iterative algorithm. We propose a deep unrolling framework for the phase retrieval problem that augments the alternating projections method with a learned denoising step. Our method obtains high-quality reconstructions without the need for retraining in settings where conventional model-based alternatives struggle, such as when the dynamic range of the measurement operator changes significantly at test time.
Electrical & Computer Engineering, University of Washington · NSF-Simons National Institute for Theory and Mathematics in Biology · Department of Computer Science, University of Chicago +3
Fourier phase encodes important spatial image structure, but recovering an image without measured spectral magnitude requires additional constraints and leaves absolute intensity ambiguous. We propose a projection-based two-stage framework that combines Fourier-phase and spatial-support constraints with an image-specific neural prior. The first stage alternates constraint enforcement with regularized neural-prior updates, while the second performs phase/support refinement alone with guaranteed convergence. We evaluate two neural-prior implementations on the same 77 microscopy images and compare them with a constraint-only baseline. After 500 final refinement passes, the best-performing variant achieves 31.41 dB pooled PSNR, 35.75 dB mean PSNR, and 0.9531 mean SSIM, improving pooled PSNR by 1.51~dB and reducing pooled MSE by 29.3% relative to the baseline. The results demonstrate the benefit of combining neural guidance with explicit constraint refinement at the evaluated iteration budget, while showing that lower phase residual alone does not guarantee greater reconstruction accuracy.
Ene Meco, Ahmet Enis Cetin
Dept. of Electrical and Computer Engineering, University of Illinois at Chicago, IL, 60607, USA