acceptodds
Under review as a conference paper at ICLR 2027

PhaseFuse: Recursive Fusion of Analytic Solvers for Self-Supervised Phase Retrieval

Abstract

Phase retrieval reconstructs an object from intensity-only measurements, but analytic solvers can respond differently to limited and noisy observations. We study whether these complementary estimates can be combined without clean reconstruction targets or natural-image training data. PhaseFuse aligns the global phases of five analytic solver outputs and of a candidate that keeps the current estimate, then uses a compact network to predict spatial fusion weights. The normalized blend is trained on procedural phase objects using independent noisy input and target measurements. At inference, a recursive schedule restarts the solvers from the fused estimate after short update blocks, reusing the single-pass-trained selector. On 200 BSDS500 test images encoded as phase-only objects, with two coded diffraction patterns and 30 dB Gaussian measurement SNR, recursive pixel-wise fusion (PhaseFuse-PS) reaches 35.97 dB mean phase-PSNR. This is 7.27 dB above single-pass fusion and 9.32 dB above the strongest configured standalone expert, at 1.76 times the measured inference time of single-pass fusion, and every test image improves. Image-level selection (PhaseFuse-GH) reaches 36.45 dB under the same condition. Routing becomes more concentrated as the snapshot count increases. These results support further study of learned solver combination and repeated constraint enforcement under matched computation and on experimental imaging data.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Phase retrieval, Self-supervised learning, Inverse problems

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_GdiuOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Different phase-retrieval solvers can produce different estimates from the same measurements. This paper asks whether combining those estimates can improve reconstruction without clean reconstruction targets or natural-image training data. PhaseFuse first aligns the solver outputs to a common global phase so they can be meaningfully combined. A small convolutional network assigns weights to five solver estimates and a sixth option that simply keeps the current estimate. Their weighted sum is then normalized to unit magnitude at each pixel, enforcing the assumption that the unknown object varies only in phase.

Training uses procedurally generated objects, each measured twice with independent noise. The first measurement drives the solvers and the selector; the second provides the target for the measurement loss. At inference, the solvers take a few steps, combine their estimates, and start again from that shared result. Repeating this process lets each solver use the updated estimate, while keeping the selector trained for single-pass fusion unchanged. The main evaluation uses 200 BSDS500 test images, with two coded diffraction patterns per image and a measurement SNR of 30 dB.

Critical strengths

The method addresses a practical question: can learning improve the use of existing reconstruction algorithms without requiring clean object targets? Restricting the network to combining analytic candidates gives the design a clear role. The phase alignment and relative-phase features also account explicitly for the global-phase ambiguity. The stated invariance argument appears sound under its nondegeneracy assumptions.

The noisy-target objective is well motivated. Under the stated independent, zero-mean target-noise assumptions, its expected measurement loss differs from the noiseless measurement loss by a parameter-independent term. The paper appropriately distinguishes this identity from a guarantee of object recovery.

The recursive improvement is substantial within the reported setup. Reusing the same selector improves every test image over single-pass fusion for the schedules evaluated in Table 2. I also appreciate the informative ablations and the candid discussion of limitations: whole-image selection can outperform spatial blending, and the three- and four-pattern cases reduce to projected, restarted TWF.

Critical weaknesses

My main concern is that the experiments do not isolate the contribution of learning from the changes to the analytic iterations. Recursive fusion repeatedly imposes the known unit-modulus constraint, restarts the solvers, and resets their internal states. Equation (8) shows that the same procedure becomes a projected, restarted single solver when routing concentrates on one expert. This is already the observed behavior with three and four diffraction patterns. At the main two-pattern setting, I would therefore like to see projected/restarted PAP and TWF with the same block schedules, together with a simple measurement-residual selector. These controls would establish how much of the reported gain comes from learned decisions. Section 5 acknowledges this issue accurately, but it remains central to assessing the contribution.

The routing comparisons also leave room for simpler alternatives. Uniform mixing includes RAF, which performs poorly under the chosen configuration, so beating that baseline does not by itself demonstrate a need for a learned selector. The retrospective PAP–TWF average reaches 35.41 dB, close to the learned spatial blend at 35.97 dB. Because that pair was chosen after inspecting test results, it should not be treated as a deployable competitor. It does, however, motivate evaluating a small fixed subset selected on disjoint procedural validation data. Moreover, whole-image hard selection and whole-image soft blending reach 36.45 and 36.09 dB, respectively. These results support adaptive selection, but give limited evidence that spatial mixing is essential under recursion. Decision-wise routing at the two-pattern setting would help explain what the selector contributes.

The computational comparison needs further clarification. An equal budget of 200 updates per expert gives the fusion system 1,000 aggregate updates across its five iterative solvers, whereas a standalone solver receives 200. These updates have different costs, so update counts alone cannot establish a matched-compute advantage. The reported 1.76x runtime increase compares recursive fusion with single-pass fusion, not with the strongest standalone solver. An accuracy-versus-runtime comparison against tuned standalone and projected/restarted baselines would be more informative. It should include residual evaluation and network overhead.

Finally, the smaller differences among routing rules and training sources need stronger uncertainty estimates. The standard deviations in Table 2 describe variation across images, while the experiments use one training seed and one noise realization per condition. Repeated runs would help determine whether the sub-dB rankings are stable. The procedural-versus-natural comparison also uses 10,000 distinct procedural objects but only 200 distinct natural images. This supports the feasibility of procedural training under the tested protocol, rather than an intrinsic advantage of procedural data. The manuscript already discloses this difference, and the conclusion should remain correspondingly narrow.

Questions for authors
  1. At two diffraction patterns and 30 dB SNR, how do projected/restarted PAP and TWF perform under the same recursive schedules? What is the remaining gain from the learned selector?
  2. Can a fixed expert subset or a measurement-residual routing rule, chosen using only procedural validation data, approach the learned results?
  3. How do the methods compare at matched wall-clock time, including all solver, residual, alignment, and network costs?
  4. At the two-pattern setting, does whole-image selection switch experts early in the trajectory, or select TWF throughout? Table 3 reports the final decision and does not resolve this.
  5. How stable are the main comparisons across independent training and noise seeds?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Gdiu

Dear Reviewer Gdiu,

We would like to thank Reviewer Gdiu very much for its exceptionally thoughtful review. Numbers are mean phase-PSNR (dB) on the 200 BSDS500 test images with two diffraction patterns (L = 2), at the paper's SNRs of 20, 30 and 40 dB, averaged over 3 training seeds.

Main concern and Question 1. Isolating the contribution of learning

These are the controls you asked for, with no network involved:

Control (L = 2)20 dB30 dB40 dB
PAP, projected/restarted every 8 updates (same 25 × 8 schedule)24.7433.8742.90
PAP, projected after every update25.8335.0244.03
TWF, projected/restarted every 8 updates (same 25 × 8 schedule)26.6236.4546.34
TWF, projected after every update27.2437.1847.16
Measurement-residual selector (per image, no training)28.3437.6147.17
PhaseFuse (learned selector)30.3338.4647.37

The controls show that imposing the unit-modulus constraint after every update is the most stable way to run both solvers. PhaseFuse is designed to fuse the best available solvers, so we updated its library to this more stable form: TWF projected after every update, plus TV-regularized versions of it at five strengths. The network, the label-free training and the procedural data are unchanged. With this library the changes to the analytic iterations are already inside every expert: restarting changes nothing (25 blocks vs one pass: −0.07 / −0.01 / 0.00 dB), so PhaseFuse uses a single fusion pass, and routing does not concentrate on one expert (normalized routing entropy about 0.67 at L = 2 and 0.56–0.64 at L = 3 and 4).

Conclusion. The contribution of learning is now isolated. At two diffraction patterns and 30 dB SNR, the remaining gain from the learned selector is +1.28 dB over the best projected/restarted solver (TWF) and +3.44 dB over PAP. Over all SNRs it is +3.09 / +1.28 / +0.21 dB over the best control and +1.99 / +0.85 / +0.20 dB over the measurement-residual selector.

Routing comparisons and Question 2. Fixed subset and residual routing

The updated library contains no RAF. We scored every equal-weight subset of its six experts, and every single expert, by the held-out measurement error on 50 disjoint procedural validation objects:

Rule chosen using only procedural validation data (L = 2)20 dB30 dB40 dBPhaseFuse better on
Best single expert (TV-TWF λ = 0.005)29.3537.3640.23200 / 200 / 200 images
Best fixed expert subset (TV-TWF at λ = 0.01, 0.005, 0.0005)29.3037.5540.80200 / 200 / 200 images
Measurement-residual routing (per image)28.3437.6147.17200 / 199 / 120 images
PhaseFuse30.3338.4647.37

Conclusion. No. The best fixed expert subset and the measurement-residual routing rule, chosen using only procedural validation data, do not approach the learned results: PhaseFuse is ahead by +0.91 to +6.57 dB and +0.20 to +1.99 dB respectively, and it is better than the fixed subset on all 200 images at every SNR.

Spatial mixing and Question 4. Whole-image selection along the trajectory

With the updated library, spatial mixing is clearly better than both whole-image rules from the same network (3 seeds):

L = 220 dB30 dB40 dB
Spatial (per-pixel) blending, PhaseFuse30.3338.4647.37
Whole-image soft blending30.1538.2447.33
Whole-image hard selection29.9938.0747.33

With this library the trajectory has a single fusion step, so whole-image selection makes one decision per image. We report it for every image (seed 0; the other seeds are similar):

L = 2Experts chosen by whole-image selection (images out of 200)
20 dBTV-TWF λ = 0.02 (129), TV-TWF λ = 0.01 (71)
30 dBTV-TWF λ = 0.002 (177), TV-TWF λ = 0.005 (23)
40 dBTV-TWF λ = 0.0005 (200)

Conclusion. Spatial mixing is now essential: it beats whole-image soft blending by +0.18 / +0.22 dB and whole-image hard selection by +0.34 / +0.39 dB at 20 / 30 dB, in every seed. Whole-image selection does not select TWF throughout: it switches between experts with the noise level, which shows that the selector contributes noise-adaptive decisions, and the per-pixel blend adds spatial adaptation on top of them.

Computational comparison and Question 3. Matched wall-clock time

Times are measured per image on an RTX 3090 (batch 50):

L = 2Time per image20 dB30 dB40 dB
TWF projected after every update, 200 updates8.8 ms27.2437.1847.16
TWF projected after every update, 5,000 updates≈ 220 ms27.2237.1847.16
Best standalone solver (TV-TWF λ = 0.005), 1,000 updates≈ 315 ms29.3537.3640.23
PhaseFuse (all solver, residual, alignment and network costs)320 ms30.3338.4647.37

Of PhaseFuse's 320 ms, the residuals, alignment and network take under 1 ms; the rest is the solvers. All solvers converge within 200 updates, so more runtime does not change their accuracy.

Conclusion. At matched wall-clock time, PhaseFuse is the most accurate method at every SNR: +3.11 / +1.28 / +0.21 dB over projected TWF with 25 times more updates, and +0.98 / +1.10 / +7.14 dB over the best standalone solver with 5 times more updates.

Uncertainty estimates and Question 5. Training and noise seeds

Across 3 independent training seeds, PhaseFuse gives 30.31–30.35, 38.46 and 47.32–47.41 dB at 20, 30 and 40 dB, and the sub-dB ranking (per-pixel > whole-image soft > whole-image hard) holds in every seed at 20 and 30 dB; the per-pixel lead over whole-image soft blending is +0.15 to +0.23 dB in every seed. Across 4 independent test-noise realizations at 30 dB, PhaseFuse gives 38.44–38.45 dB and its lead over projected TWF stays at +1.27 dB. All differences are paired over the same 200 images, and we will report them with 95% confidence intervals. On the procedural-versus-natural comparison, we agree and will state only the feasibility of procedural training under the tested protocol.

Conclusion. The main comparisons are stable: across training and noise seeds PhaseFuse varies by at most 0.09 dB, and every sub-dB ranking keeps its sign at 20 and 30 dB. The procedural-data conclusion will remain correspondingly narrow.

We would like to thank Reviewer Gdiu again for its generous and rigorous review. We would be very grateful if you would consider these results in your final assessment.

AuthorsOpenReview ↗

Revised paper

Revised paper

We thank Reviewer Gdiu again for the thorough review. We have uploaded the revised paper, with all changes marked in blue.

Main concern and Question 1 (contribution of learning). Appendix D.2 and Table 19 run PAP and TWF with the projection every 8 updates and after every update, together with the measurement-residual selector. Table 3 in Section 4.4 compares PhaseFuse with projected TWF and TV-TWF for one to four diffraction patterns.

Question 2 (fixed subset and residual routing). Appendix D.3 and Table 20 report the single solver, the solver combination, and the residual routing chosen using only procedural validation data, now for two to four diffraction patterns.

Question 3 (matched wall-clock time). Appendix D.5 and Table 23 report the measured time per image, including all solver, residual, alignment, and network costs.

Question 4 (whole-image selection). Appendix D.4 compares spatial blending with whole-image soft blending and whole-image selection (Table 21) and reports the solver selected for every image (Table 22).

Question 5 (seeds). Table 3 reports the mean of three training seeds with paired 95% confidence intervals, and Appendix D.8 adds training seeds, new noise realizations, and unseen masks. The claim on procedural against natural training data is narrowed to the feasibility of procedural training in Sections 1 and 4.6.

Reviewer_hbwzOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes PhaseFuse, which learns to combine analytic solvers for phase retrieval. It aligns their phase estimates before selecting or blending their outputs. The selector is trained on procedural images using independently corrupted measurements, without clean reconstruction targets. At inference, the solvers repeatedly restart from the combined estimate. On the main simulated benchmark, recursive fusion improves phase-PSNR by 9.32 dB over the strongest configured standalone solver.

Critical strengths

-Training without clean targets or natural images is useful for imaging problems where suitable training data are scarce.

-The ablations are informative, and the authors clearly acknowledge where the results favor simple solver scheduling over learned fusion.

Critical weaknesses

-The benefit of learning is not sufficiently isolated. With three or four masks, the method effectively becomes restarted TWF with repeated projection. At two masks, a simple PAP–TWF average nearly matches learned fusion, although that pair was selected after inspecting test results.

-Equal updates per solver do not mean equal total computation. The budget studies help, but comparisons with tuned, restarted single solvers at matched runtime are still missing.

-Evaluation uses simulated phase-only objects, fixed masks shared between training and testing, and one training seed. This leaves robustness beyond the tested setup uncertain.

Questions for authors

-How much gain remains over restarted, projected TWF and a validation-selected solver combination at matched runtime?

-Does the trained selector retain its performance on unseen masks and across training seeds?

5 comments
AuthorsOpenReview ↗

Response to Reviewer hbwz

Dear Reviewer hbwz,

We would like to thank Reviewer hbwz very much for its clear and fair review. Numbers are mean phase-PSNR (dB) on the 200 BSDS500 test images with two patterns (L = 2) unless stated otherwise, at the paper's SNRs of 20, 30 and 40 dB, averaged over 3 training seeds.

PhaseFuse is designed to fuse the best available analytic solvers. To isolate the benefit of learning, we updated its library to more stable solvers that already include the repeated projection: TWF with the unit-modulus projection after every update, plus TV-regularized versions of it at five strengths. The network, the label-free training and the procedural data are unchanged. Because these solvers already project at every update, and restarting them changes nothing (25 blocks vs one fusion pass: −0.07 / −0.01 / 0.00 dB), any gain that remains over them comes from learning. PhaseFuse therefore uses a single fusion pass.

Weakness 1. Isolating the benefit of learning

With three or four masks, the updated library no longer collapses onto one solver: the normalized routing entropy at the end of training is 0.56–0.64 in all 6 runs (0 means all weight on one solver). Compared directly with TWF with repeated projection:

LProjected TWF (20 / 30 / 40 dB)PhaseFuse (20 / 30 / 40 dB)
330.23 / 40.10 / 50.0632.32 / 40.91 / 50.10
431.93 / 41.78 / 51.7533.64 / 42.41 / 51.65

At two masks, we replaced the PAP–TWF pair, which was selected after inspecting test results, with a solver combination selected without any test data (Question 1 below).

Conclusion. The benefit of learning is isolated: the solvers already include the repeated projection, the method no longer reduces to restarted TWF at three or four masks, and PhaseFuse stays above or within 0.1 dB of projected TWF in every setting, with gains of up to +2.09 dB. This results will be added in the respective table.

Weakness 2 and Question 1. Matched runtime

For the validation-selected solver combination, we scored every equal-weight subset of the six solvers by the held-out measurement error on 50 procedural validation objects (no ground truth). All times are measured per image on an RTX 3090 (batch 50) and include every cost:

L = 2Time per image20 dB30 dB40 dB
Projected TWF, 5,000 updates≈ 220 ms27.2237.1847.16
Restarted, projected TWF (25 blocks of 8 updates)≤ 9 ms26.6236.4546.34
Best single solver on validation (TV-TWF λ = 0.005), 1,000 updates≈ 315 ms29.3537.3640.23
Validation-selected solver combination (TV-TWF at λ = 0.01, 0.005, 0.0005, equal weights)≈ 190 ms29.3037.5540.80
PhaseFuse320 ms30.3338.4647.37

All solvers converge within 200 updates, so 1,000 or 5,000 updates, or an exponentially decaying step size, change projected TWF by at most 0.01 dB. Giving the baselines PhaseFuse's runtime therefore does not change their results.

Conclusion. At matched runtime, the gain that remains is +3.11 / +1.28 / +0.21 dB over restarted, projected TWF in its strongest form (projected after every update, 5,000 updates), and +1.03 / +0.91 / +6.57 dB over the validation-selected solver combination, with PhaseFuse better than that combination on all 200 images at every SNR. We will add this important result in the ablations supplementary.

Weakness 3 and Question 2. Unseen masks, training seeds and phase-only objects

Unseen masks: on three random mask sets never seen in training (L = 2, 30 dB, no retraining), PhaseFuse gives 38.44 / 38.46 / 38.46 dB, against 38.46 dB on the training masks, and its lead over projected TWF stays at +1.27 dB on each set.

Training seeds: across 3 independent training seeds, PhaseFuse gives 30.31–30.35, 38.46 and 47.32–47.41 dB at 20, 30 and 40 dB. Across 4 new test-noise realizations at 30 dB, it gives 38.44–38.45 dB.

Phase-only objects: we added an experiment on experimental optical data, CXIDB-175 (Levitan et al., 2020), whose sample is an amplitude object (a Siemens star) with a probe calibrated from a ptychographic scan. With a library of solvers suited to that setup and the same network trained only on synthetic objects, PhaseFuse reaches 21.95 dB magnitude PSNR on 10 held-out frames. This is above the best single solver of that library chosen without labels (21.35 dB when chosen on synthetic validation, 21.82 dB when chosen on the real frames' own held-out measurements).

Conclusion. Yes. The trained selector retains its performance on unseen masks (within 0.02 dB) and across training seeds (within 0.09 dB), and the method also holds on experimental data with a non-phase-only object. We will add this ablation on real data on the main paper.

We thank Reviewer hbwz again for its careful and balanced review. We would be very grateful if you would consider these results in your final assessment.

References

Abraham L. Levitan, Kahraman Keskinbora, Umut T. Sanli, Markus Weigand, and Riccardo Comin. Single-frame far-field diffractive imaging with randomized illumination. Optics Express, 28(25):37103–37117, 2020. https://doi.org/10.1364/OE.397421. Dataset: CXIDB ID 175, https://doi.org/10.11577/1722944

Thanks for the additional experiments!

The unseen-mask and seed results address my robustness questions, and the validation-selected comparisons strengthen the case for learned fusion at two masks.

For three and four masks, could you also compare against a validation-selected baseline from the updated library? The comparison with plain projected TWF alone leaves some uncertainty about how much of the gain comes from learning versus TV regularization.

Also, why is the 200-update restarted TWF baseline 0.60–0.82 dB below the 5,000-update version if convergence occurs within 200 updates? Do these runs use different settings?

Additionally, please make the updated solver library and single-pass setup clear in the revised paper.

Reviewer_ScZJOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents a self-supervised phase retrieval (PR) method that fuses the outputs of different analytic solvers. During training, the paper uses fractals to generate synthetic training images, and it trains a solver selector for pixel-level fusion from two independent noisy measurements of each image, without clean targets. This gives the single-pass PhaseFuse-PS. After training, the paper applies the same selector recursively, without retraining: the solvers run in short blocks on the same measurement and restart from the fused estimate after each block. The recursive approach is tested with the pixel-level PhaseFuse-PS and with the image-level PhaseFuse-GH, which picks one solver for the whole image, under different SNRs, image sizes and numbers of snapshots, and compared with the individual solvers. In these simulated experiments on phase-only objects, single-pass PhaseFuse-PS improves over the best individual solver with one or two snapshots, and the recursive versions of both methods improve the reconstruction substantially (by about 9–10 dB over the best solver with two snapshots at 30 dB SNR). This could be helpful for data-limited inverse problems involving PR. However, several technical details and design choices remain to be clarified, as outlined in the weaknesses below.

Critical strengths
  1. The method learns to combine the outputs of six expert solvers without paired clean data. The six experts are five analytic phase retrieval (PR) solvers plus one expert that keeps the current estimate. A compact network predicts per-pixel weights over the experts. It is trained only on pairs of independent noisy measurements of each object, following the Noise2Noise principle.

  2. The training objects are procedurally generated fractal images, following the one-instance FractalDB approach. This removes the need for natural-image training data and provides diverse image structures. Under recursive fusion, the network trained this way slightly outperforms one trained on natural images, both on BSDS500 and on unseen FFHQ images.

Critical weaknesses
  1. The paper does not include a loss ablation study. It should evaluate the contribution of each loss term (L_meas, L_rot and L_TV) and explain how each one affects the final reconstruction performance. There is no run without L_rot or L_TV, no sensitivity study for λ_rot and λ_TV, and no analysis of how either term affects training or the recursive inference results. For a self-supervised method, the loss design is central.
  2. The method is not compared with existing self-supervised or unsupervised PR methods, such as Chen et al. (2022; DOI: 10.1109/TSP.2022.3170710) and Quan et al. (2023; DOI: 10.1609/aaai.v37i2.25306). The paper itself states that it does not compare with learned PR methods.
  3. The setup of PhaseFuse-GH is unclear. The selector is trained only with the per-pixel rule (PhaseFuse-PS), so the paper should explain more explicitly how the image-level selection is applied at inference, which solver it selects at each recursive step, and how it differs from PhaseFuse-PS.
  4. The presentation is difficult to follow as there is too much notation, and the logic sometimes jumps between ideas without sufficient explanation. Some symbols, such as the residual ρ_e, appear only once, and the same letter is later reused for other quantities (e.g., ρ for the ADMM penalty), which makes their roles unclear.
Questions for authors
  1. Could the authors ablate the two loss terms that do not use the measurements, L_rot and L_TV? L_TV favors smooth reconstructions and may itself suppress noise, so part of the reported gain could come from this smoothing rather than from fusion. How do the results change during training and at inference when each term is removed?

  2. How exactly is PhaseFuse-GH applied at inference? Why does it outperform PS in several settings when the network is trained only for PS? What determines which solver GH selects at each recursive step? Please report the solver chosen at every step, not only at the final one.

  3. The paper mixes the terms “selection” and “fusion.” The per-pixel rule (PS) is a fusion: each output pixel is a weighted blend of all experts. The introduction even says the method combines solvers “rather than select one.” Yet the network is called a “selector” throughout, including in Figure 1, which suggests that one expert’s output is chosen. Could the authors use these terms consistently and clarify which rule is the main method?

  4. Under the current test settings, fractal training works about as well as training on natural images for PhaseFuse-PS (Sec. 4.5). However, the truly data-limited case has not been tested; the smallest training sets contain 200 natural images and 1,000 procedural images. Could the current method genuinely support few-shot training with only a few training objects, or even zero-shot selection with no training at all (e.g., choosing the solver with the lowest measurement residual)?

5 comments
AuthorsOpenReview ↗

Response to Reviewer ScZJ

Dear Reviewer ScZJ,

We want to thank reviewer ScZJ for reading our paper so carefully and for the clear summary. The questions on the loss, on PhaseFuse-GH and on the terminology helped us improve the paper. One update affects all numbers below. PhaseFuse is designed to fuse the best available analytic solvers, and following the reviews we updated its library to more stable solvers: TWF with the unit-modulus projection after every update, plus TV-regularized versions of it at five strengths. The network, the label-free training and the procedural data are unchanged. With this library the recursive schedule gives the same accuracy as one fusion pass (25 blocks vs one pass: −0.07 / −0.01 / 0.00 dB), so we report the single pass. Numbers are mean phase-PSNR (dB) on the 200 BSDS500 test images with two patterns (L = 2), at the paper's SNRs of 20, 30 and 40 dB.

Weakness 1 and Question 1. Loss ablation

L_meas is the only term that uses the measurements, so the network cannot be trained without it. We removed each of the other two terms and varied both weights over two orders of magnitude, retraining the network each time (seed 0, everything else identical):

Loss20 dB30 dB40 dB
L_meas only (no L_rot, no L_TV)30.3538.4447.29
L_meas + L_rot (λ_rot = 0.1), updated method30.3538.4647.41
L_meas + L_rot (λ_rot = 0.05)30.3638.4247.20
L_meas + L_rot (λ_rot = 0.2)30.3638.4547.35
+ L_TV (λ_TV = 1e-4)30.3638.4547.33
+ L_TV (λ_TV = 1e-3, submitted)30.3538.4647.30
+ L_TV (λ_TV = 1e-2)30.4238.1644.03

At inference, L_TV does not help at 20 and 30 dB, and at the largest weight it over-smooths at high SNR (−0.30 dB at 30 dB, −3.38 dB at 40 dB), so the updated method uses no L_TV. Removing L_rot changes the results by at most 0.12 dB, and varying λ_rot by at most 0.21 dB; we keep it at λ_rot = 0.1. During training, every setting converges smoothly, and its label-free validation monitor reaches its best value at the final epoch, except λ_TV = 1e-2, whose best epoch is 15. On whether "part of the reported gain could come from this smoothing rather than from fusion": with L_meas alone, PhaseFuse reaches 30.35 / 38.44 / 47.29 dB, above the best single solver of its library chosen without labels (29.35 / 37.36 / 40.23 dB) by +1.00 / +1.08 / +7.06 dB.

Conclusion. The gain does not come from the smoothing of L_TV. Removing L_TV costs nothing, so the updated method does not use it; L_rot has only a minor effect; and with the measurement loss alone PhaseFuse remains clearly above the best single solver. The gain comes from fusion. We will add this resuts as ablation in the main paper for its important analysis.

Weakness 2. Comparison with self-supervised or unsupervised PR methods

We ran both with their official code, changing only the measurement operator and the noise model. Quan et al. (2023) was trained with its own self-supervised loss on our 10,000 training measurements, for the same 20 epochs as our network. Chen et al. (2022) (DeepMMSE) optimizes an untrained network for each image (100,000 iterations), so we ran it on the first 12 test images and report PhaseFuse on the same images.

L = 2Test images20 dB30 dB40 dBTime per image
Quan et al. (2023)20027.4631.1532.29one network pass
PhaseFuse20030.3338.4647.37320 ms
Chen et al. (2022), DeepMMSEfirst 1231.1033.1533.36≈ 4 min
PhaseFusesame 1230.2338.2647.22320 ms

Conclusion. PhaseFuse is more accurate than Quan et al. (2023) at every SNR (+2.87 / +7.31 / +15.08 dB) and than Chen et al. (2022) at 30 and 40 dB (+5.11 / +13.86 dB, on all 12 images). At 20 dB it is statistically tied with Chen et al. (95% CI of the difference [−1.84, +0.12] dB) while being about 780 times faster. We will add this resuts as ablation in the main paper.

Weakness 3 and Question 2. PhaseFuse-GH

How GH is applied at inference and how it differs from PS: GH uses the same trained network as PS. The network outputs one logit per solver at every pixel. PS turns these logits into per-pixel softmax weights and blends the phase-aligned solver outputs. GH instead averages each solver's logits over the whole image and returns the single solver with the largest average, so it makes one decision per image.

Why GH outperformed PS in several settings: GH wins when one solver is better than all the others over the whole image. In the submitted library this was TWF in several settings. With the updated library, PS is ahead at every SNR at L = 2 (3 seeds), and GH stays ahead only in the noisiest single-pattern case (L = 1, 20 dB, +0.34 dB), where it picks the strongest TV solver for every image:

L = 220 dB30 dB40 dB
PhaseFuse-PS (per-pixel blend)30.3338.4647.37
Image-level soft weights30.1538.2447.33
PhaseFuse-GH (one solver per image)29.9938.0747.33

What determines the solver and which one is chosen at each step: with the updated library, PhaseFuse makes a single fusion step, so the table below reports the solver GH chooses at that step, for every image. The choice is set by the image-averaged logits, and it follows the noise level (seed 0; the other seeds are similar):

L = 2Solver chosen by GH (number of images out of 200)
20 dBTV-TWF λ = 0.02 (129), TV-TWF λ = 0.01 (71)
30 dBTV-TWF λ = 0.002 (177), TV-TWF λ = 0.005 (23)
40 dBTV-TWF λ = 0.0005 (200)

Conclusion. GH is the same trained network read out differently: one solver per image instead of a per-pixel blend. It can beat PS only when a single solver is best over the whole image. With the updated library, PS is better at every SNR at L = 2, so PS is the main method and GH becomes an ablation. We will add this description and these tables to the paper.

Question 3 and Weakness 4. Terminology and notation

We agree with both points. The per-pixel blend (PhaseFuse-PS) is the main method, and we will call the network the "fusion network" throughout, including Figure 1; GH becomes an ablation named "image-level selection". The ADMM solver is no longer in the library, so ρ will denote only the solver residual. We will add a notation table, remove the symbols used only once, and organize the method section in the order of the pipeline: solvers, phase alignment, fusion network, training loss, inference.

Conclusion. "Fusion" will be used consistently, PS is the main method, and every symbol will have a single meaning.

Question 4. Few-shot and zero-shot

We retrained the network on 100, 10 and 1 procedural objects, with the same 1,580 training steps (seed 0). For zero-shot selection we used exactly your suggestion: for each image, the solver with the lowest measurement residual.

Training objects20 dB30 dB40 dB
10,00030.3538.4647.41
10030.3438.3747.13
1030.1738.3146.89
127.5537.6547.33
0 (zero-shot: lowest-residual solver)28.3437.6147.17

With one object, the network puts all its weight on a single solver; that solver happens to be the best one at 40 dB, but the network no longer adapts to the noise level.

Conclusion. Yes, PhaseFuse genuinely supports few-shot training: 100 procedural objects match full training at 20 and 30 dB, and 10 objects stay within 0.2 dB at those SNRs. Zero-shot selection by the lowest measurement residual is clearly weaker (about 2 dB lower at 20 dB), so a small amount of label-free training is what makes the fusion work.

We would like to thank again Reviewer ScZJ for its thorough and constructive review. We would be very grateful if you would consider these results in your final assessment.

References

Mingqin Chen, Peikang Lin, Yuhui Quan, Tongyao Pang, and Hui Ji. Unsupervised phase retrieval using deep approximate MMSE estimation. IEEE Transactions on Signal Processing, 70:2239–2252, 2022. https://doi.org/10.1109/TSP.2022.3170710

Yuhui Quan, Zhile Chen, Tongyao Pang, and Hui Ji. Unsupervised deep learning for phase retrieval via teacher-student distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):2128–2136, 2023. https://doi.org/10.1609/aaai.v37i2.25306

AuthorsOpenReview ↗

Revised paper

Revised paper

We thank Reviewer ScZJ again for the careful review. We have uploaded the revised paper, with all changes marked in blue.

Weakness 1 and Question 1 (loss ablation). Appendix D.6 and Table 24 remove L_rot and L_TV and vary both weights, retraining the network each time. Section 4.4 states that, with the projected library, PhaseFuse uses no L_TV.

Weakness 2 (self-supervised PR methods). Appendix D.9 and Table 28 compare PhaseFuse with DeepMMSE (Chen et al., 2022) and Quan et al. (2023), together with other learned PR methods, and Section 5 now describes these comparisons.

Weakness 3 and Question 2 (PhaseFuse-GH). Section 3.3 now defines PhaseFuse-GH as image-level selection: the same trained network read out with one decision per image, reported as an ablation. Appendix D.4 compares it with the per-pixel blend (Table 21) and reports the solver it chooses for every image (Table 22).

Question 3 and Weakness 4 (terminology and notation). The network is called the fusion network throughout, including Figure 1, the method section follows the order of the pipeline (Sections 3.1 to 3.6), and Appendix A.1 (Table 5) lists the notation.

Question 4 (few-shot and zero-shot). Appendix D.7 and Table 25 retrain the network on 100, 10, and 1 procedural objects and report zero-shot selection by the measurement residual.

Reviewer_WySjOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The manuscript presents PhaseFuse, a self-supervised framework for phase retrieval (PR) that learns to spatially combine candidate reconstructions generated by a diverse library of fixed analytic solvers (Parallel Alternating Projection [PAP], Relaxed Averaged Alternating Reflections [RAAR], Truncated Wirtinger Flow [TWF], Reweighted Amplitude Flow [RAF], and Least Absolute Deviation ADMM [LAD-ADMM]) alongside a spectral initialization expert.

PhaseFuse addresses two primary challenges in computational imaging: (1) combining complementary non-convex PR solvers without clean ground-truth targets or natural-image priors, and (2) preventing solver stagnation under noisy, limited-snapshot regimes.

On 200 BSDS500 test images (128 X128 phase-only objects, L=2 coded diffraction patterns [CDP], 30 dB SNR), recursive fusion achieves 35.97 dB (PhaseFuse-PS) and 36.45 dB (PhaseFuse-GH) mean phase-PSNR—outperforming single-pass fusion (28.70 dB) by 7.27 dB and the best standalone expert (TWF, 26.65 dB) by 9.32 dB at 1.76X single-pass inference wall-clock time.

Critical strengths
  1. Rigorous Global Phase Alignment Mechanics: The explicit phase alignment step before spatial weighting is mathematically sound and essential. The authors rigorously prove global phase invariance for the selector input and global phase equivariance for the combined output.

  2. Target-Free & Data-Agnostic Self-Supervised Pipeline: Coupling intensity-domain Noise2Noise loss formulation with procedural IFS fractal object generation avoids dataset leakage, natural image priors, and supervised clean targets. Demonstrating that a selector trained on abstract fractals transfers seamlessly to natural images (BSDS500) and human faces (FFHQ) is a notable result.

  3. Substantial Performance Gains via Recursive Scheduling: Moving from single-pass combination to recursive update blocks yields a dramatic +7.27 dB improvement at the exact same per-solver iteration budget (N=200). The variance across test images drops, showing high consistency.

  4. Comprehensive Ablations & Negative Results Analysis: The manuscript includes thorough ablations covering per-pixel vs. image-level routing, softmax temperature sweeps, procedural vs. natural training data, expert budget calibration, and the failure of RL/REINFORCE policy gradient optimization.

Critical weaknesses
  1. Confounding Multi-Expert Fusion with Single-Expert Recursive Restarts (The "TWF Collapse") The Argument: The paper attributes its +9.32 dB gain primarily to "learned solver combination" and "recursive expert fusion" across heterogeneous algorithms. Why it is Weak:As shown in Table 3, when the snapshot count increases to L=3 or L=4, the learned selector assigns >0.99999 weight to TWF across all pixels, decisions, and images. At L >= 3, PhaseFuse performs zero spatial mixing or expert combination; it collapses entirely into running TWF for τ\tau steps, projecting onto the unit circle, and restarting TWF for D blocks.

  2. Asymmetric Computational Budgets & Missing Equal-FLOP Baselines The Argument: The paper claims recursive fusion is evaluated at "an identical budget of 200 updates per solver". Why it is Weak:Comparing PhaseFuse (which executes 5 iterative analytic experts for 200 iterations each = 1,000 total solver iterations, plus 1 spectral expert, plus D=25 U-Net forward passes, plus candidate phase alignments and residual calculations) against a single standalone expert running 200 iterations is an asymmetrical compute comparison. A standalone expert (e.g., TWF or PAP) allocated an equivalent computational budget of 1,000 to 5,000 continuous iterations (or run with step-size decay) may perform significantly better than its 200-iteration baseline.

  3. Oversimplified Physical Models & Forward Constraints The Argument: The formulation is evaluated exclusively on synthetic phase-only objects under additive Gaussian noise on intensity measurements with perfectly known CDP masks. Why it is Weak: Real-world coherent diffraction imaging (CDI), ptychography, and crystallography involve general complex-valued objects with non-uniform, spatially varying amplitudes. Realistic optical/X-ray detector noise follows Poisson or mixed Poisson-Gaussian distributions, rather than pure additive Gaussian noise on squared magnitudes.

  4. Absence of Modern Self-Supervised & Deep Prior PR Baselines The Argument: PhaseFuse is evaluated almost exclusively against its own constituent analytic sub-experts. Why it is Weak: The literature contains several physics-informed, untrained, or self-supervised phase retrieval frameworks that do not require clean targets, such as PhysenNet, AutoPhaseNN, PtychoPINN, prDeep, or Deep Image Prior for PR. Without comparing against DIP-PR or self-supervised unrolled networks, it is unclear whether fusing classical solvers offers any advantage over optimizing an implicit neural representation directly from measurements.

Questions for authors

Listed above.

4 comments
AuthorsOpenReview ↗

Response to Reviewer WySj

Dear Reviewer WySj,

We would like to thank very much Reviewer WySj for its careful and detailed review. Unless stated otherwise, numbers are mean phase-PSNR (dB) on the 200 BSDS500 test images (128 × 128 phase-only objects), with L = 2 coded diffraction patterns, at the paper's SNRs of 20, 30 and 40 dB, averaged over 3 training seeds.

W1. Confounding Multi-Expert Fusion with Single-Expert Recursive Restarts (The "TWF Collapse")

PhaseFuse is designed to take the best available analytic solvers as its library and to learn, without clean targets, how to combine them. In the submitted library, TWF was clearly the strongest solver with three or four patterns, so the selector learned to send all weight to it. To test "learned solver combination" when no single solver dominates, we updated PhaseFuse to a more stable library: TWF with the unit-circle projection after every update (the procedure you describe, now used as an expert). The rest of PhaseFuse is unchanged: the same network, the same label-free measurement loss and the same 10,000 procedural training objects, retrained for each L with 3 seeds. Each expert runs 200 updates from the shared initialization. With this library the selector no longer collapses. At L = 3 and L = 4, the normalized routing entropy at the end of training is 0.56–0.64 in all 6 runs (0 means all weight on one expert, 1 means uniform weights). Averaged over L = 2–4 and the three SNRs, the experts receive 25.7%, 19.7%, 19.1%, 13.1%, 12.8% and 9.5% of the fusion weight, and the most chosen expert never takes more than 52% in any setting, against more than 99.999% for TWF in the submitted library. The fusion is also spatial: at L = 3 and L = 4, per-pixel weights beat a single weight per image from the same network by 0.17–0.27 dB at 20 and 30 dB.

LProjected TWF alone (20 / 30 / 40 dB)PhaseFuse (20 / 30 / 40 dB)Gain (dB)
227.24 / 37.18 / 47.1630.33 / 38.46 / 47.37+3.09 / +1.28 / +0.21
330.23 / 40.10 / 50.0632.32 / 40.91 / 50.10+2.09 / +0.81 / +0.04
431.93 / 41.78 / 51.7533.64 / 42.41 / 51.65+1.71 / +0.63 / −0.10

On "restarting TWF for D blocks": with the updated library, the recursive schedule (D = 25 blocks of 8 updates) gives the same accuracy as one fusion pass (30.25 / 38.45 / 47.37 vs 30.33 / 38.46 / 47.37 dB), because every expert already enforces the unit-circle constraint at each update. We therefore use the single pass (D = 1), and restarts play no role in the gains above.

Conclusion. The gain comes from learned solver combination, not from restarting one solver. With a stable library in which no expert dominates, PhaseFuse mixes the experts spatially at every L, and it is above or within 0.1 dB of the procedure you describe in every setting, with gains of up to +3.09 dB where the measurements are most limited. We will replace Table 3 with this analysis.

W2. Asymmetric Computational Budgets & Missing Equal-FLOP Baselines

We agree that iteration counts do not measure compute, so we now report measured wall-clock time with every cost included. With the updated library, PhaseFuse runs its 6 experts for 200 iterations each and then one network pass (D = 1). On an RTX 3090 (batch 50), the full pipeline (all experts, candidate phase alignments, residual calculations and the network) takes 320 ms per image, of which the alignments, residuals and network take under 1 ms. We gave standalone TWF and PAP equivalent and larger budgets:

L = 2Iterations per solverNetwork passesTime per image20 dB30 dB40 dB
PhaseFuse (D = 1)6 × 2001320 ms30.3338.4647.37
PhaseFuse (D = 25)6 × 20025> 320 ms30.2538.4547.37
Projected TWF1,0000≈ 44 ms27.2337.1847.16
Projected TWF5,0000≈ 220 ms27.2237.1847.16
Projected TWF + step-size decay1,0000≈ 44 ms27.2437.1847.16
Projected PAP2000< 9 ms25.8335.0244.03

Both standalone solvers have converged by 200 iterations, so 1,000 or 5,000 continuous iterations leave them unchanged; at 30 dB, PAP stays at 35.02 dB from 200 to 2,000 iterations. For step-size decay, the TWF step decays exponentially to 30%, 10% or 1% of its initial value, and all three give the same accuracy (27.23–27.24 / 37.18 / 47.16 dB). PAP has no step size. The same convergence holds for every expert in the library.

Conclusion. A standalone expert does not perform better with 1,000 to 5,000 continuous iterations or with step-size decay. At comparable measured runtime (220 ms vs 320 ms), PhaseFuse stays ahead of the strongest standalone solver by +3.11 / +1.28 / +0.21 dB, and more iterations do not close this gap. We will add this ablation on supplementary as is a very important analysis.

W3. Oversimplified Physical Models & Forward Constraints

On the noise: we tested Poisson and mixed Poisson-Gaussian noise (half of the noise power from each) at the same total SNR. The network was trained with Gaussian noise only and was not retrained:

L = 2, test noiseProjected TWF (20 / 30 dB)PhaseFuse (20 / 30 dB)
Poisson30.20 / 40.2532.59 / 41.12
Mixed Poisson-Gaussian28.47 / 38.3931.24 / 39.49

On the object and the real forward model: we tested PhaseFuse on experimental data, CXIDB-175 (Levitan et al., 2020). It is optical far-field diffraction with a randomized-phase probe, and the sample is a chrome-on-glass Siemens star, an object with non-uniform, spatially varying amplitude and flat phase. The probe is not perfectly known: it was calibrated from a ptychographic scan. Each of 10 held-out frames is reconstructed from its single diffraction pattern and scored by magnitude PSNR against the ptychographic reconstruction. Here PhaseFuse uses a different library, the solvers suited to this setup (amplitude flow and randomized probe imaging, with and without TV regularization), with a non-negative magnitude constraint instead of unit modulus. The same network is trained only on synthetic objects simulated through this forward model, with the measurement loss. PhaseFuse reaches 21.95 dB. The best single solver of this library reaches 21.35 dB when chosen on the same synthetic validation data (+0.60 dB, better on 10/10 frames) and 21.82 dB when chosen from the real frames' own held-out measurements (+0.13 dB, 8/10 frames).

Conclusion. PhaseFuse keeps its advantage under Poisson and mixed Poisson-Gaussian noise without retraining. On experimental data, with a spatially varying amplitude (another study regime), a calibrated probe and a completely different solver library, it again beats the best solver of its library. Objects with both varying amplitude and varying phase remain untested, and we will state this as a limitation.

W4. Absence of Modern Self-Supervised & Deep Prior PR Baselines

Our proposed method is an Unsupervised training method, thats why we didn't use regular supervised methods for the experiments, however based on already pretrained public weigths, we ran every method you name with its official code, changing only what our CDP measurement model requires. For DIP-PR we used DeepMMSE (Chen et al., 2022); for self-supervised unrolled networks, Quan et al. (2023).

L = 2, 200 images20 dB30 dB40 dBTime per image
prDeep (Metzler et al., 2018)30.2937.1642.53≈ 6.5 s
Quan et al. (2023), self-supervised unrolled27.4631.1532.29one network pass
AutoPhaseNN (Yao et al., 2022), with its 50 ER steps11.6513.4814.821.3 ms
PtychoPINN (Hoidn et al., 2023)6.416.416.410.2 ms
PhaseFuse30.3338.4647.37320 ms
First 12 images (per-image optimization)
DeepMMSE (Chen et al., 2022), DIP-PR31.1033.1533.36≈ 4 min
PhysenNet (Wang et al., 2020)20.2322.9122.95≈ 1 min
PhaseFuse (same 12 images)30.2338.2647.22320 ms

prDeep uses the authors' solver and pretrained DnCNN denoisers (Zhang et al., 2017) on the real and imaginary parts. AutoPhaseNN and PtychoPINN keep their architectures and losses (2D, our CDP physics layer), trained on our 10,000 measurements with their papers' schedules (70 and 60 epochs; 20 epochs gave similar results). They did not learn this inversion, as they were designed for a compact support or overlapping scans, which random-mask CDP lacks. We will release all ports.

Conclusion. Fusing classical solvers does offer an advantage over optimizing an implicit neural representation directly from measurements: PhaseFuse is the most accurate method at 30 and 40 dB, ties prDeep and DeepMMSE at 20 dB (95% CIs include zero), and is about 20 and 780 times faster than them.

We would like to thank again Reviewer WySj for the generous review. We would be very grateful if you would consider these results in your final assessment.

References

Mingqin Chen, Peikang Lin, Yuhui Quan, Tongyao Pang, and Hui Ji. Unsupervised phase retrieval using deep approximate MMSE estimation. IEEE Transactions on Signal Processing, 70:2239–2252, 2022. https://doi.org/10.1109/TSP.2022.3170710

Oliver Hoidn, Aashwin Ananda Mishra, and Apurva Mehta. Physics constrained unsupervised deep learning for rapid, high resolution scanning coherent diffraction reconstruction. Scientific Reports, 13:22789, 2023. https://doi.org/10.1038/s41598-023-48351-7

Abraham L. Levitan, Kahraman Keskinbora, Umut T. Sanli, Markus Weigand, and Riccardo Comin. Single-frame far-field diffractive imaging with randomized illumination. Optics Express, 28(25):37103–37117, 2020. https://doi.org/10.1364/OE.397421. Dataset: CXIDB ID 175, https://doi.org/10.11577/1722944

Christopher Metzler, Phillip Schniter, Ashok Veeraraghavan, and Richard Baraniuk. prDeep: Robust phase retrieval with a flexible deep network. In Proceedings of the 35th International Conference on Machine Learning, volume 80 of Proceedings of Machine Learning Research, pp. 3501–3510, 2018. https://proceedings.mlr.press/v80/metzler18a.html

Yuhui Quan, Zhile Chen, Tongyao Pang, and Hui Ji. Unsupervised deep learning for phase retrieval via teacher-student distillation. In Proceedings of the AAAI Conference on Artificial Intelligence, 37(2):2128–2136, 2023. https://doi.org/10.1609/aaai.v37i2.25306

Fei Wang, Yaoming Bian, Haichao Wang, Meng Lyu, Giancarlo Pedrini, Wolfgang Osten, George Barbastathis, and Guohai Situ. Phase imaging with an untrained neural network. Light: Science & Applications, 9:77, 2020. https://doi.org/10.1038/s41377-020-0302-3

Yudong Yao, Henry Chan, Subramanian Sankaranarayanan, Prasanna Balaprakash, Ross J. Harder, and Mathew J. Cherukara. AutoPhaseNN: Unsupervised physics-aware deep learning of 3D nanoscale Bragg coherent diffraction imaging. npj Computational Materials, 8:124, 2022. https://doi.org/10.1038/s41524-022-00803-w

Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142–3155, 2017. https://doi.org/10.1109/TIP.2017.2662206

AuthorsOpenReview ↗

Revised paper

Revised paper

We thank Reviewer WySj again for the detailed review. We have uploaded the revised paper, with all changes marked in blue.

W1 (TWF collapse). Table 3 in Section 4.4 now compares PhaseFuse, with the library of projected solvers, against projected TWF and TV-TWF for one to four masks, with paired 95% confidence intervals and the routing entropy at the end of training. The weight that each solver receives is reported in Appendix D.1, Figures 2 and 3 show this version of PhaseFuse, and the original figures are kept in Appendix C.11.

W2 (computational budgets). Appendix D.5 and Table 23 report the measured wall-clock time with every cost included, and compare PhaseFuse with projected TWF run for 1,000 and 5,000 updates and with a decaying step size.

W3 (physical models). Appendix D.8 (Tables 26 and 27) adds Poisson and mixed Poisson-Gaussian noise, and Appendix D.10 (Table 29 and Figure 8) adds the experimental optical data, CXIDB-175, with its full setup. Section 5 states that objects with both varying amplitude and varying phase remain untested.

W4 (learned baselines). Appendix D.9 and Table 28 compare PhaseFuse with prDeep, Quan et al. (2023), AutoPhaseNN, PtychoPINN, DeepMMSE, and PhysenNet, and the reproducibility statement says that our ports will be released.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.