- 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
- 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?
- Can a fixed expert subset or a measurement-residual routing rule, chosen using only procedural validation data, approach the learned results?
- How do the methods compare at matched wall-clock time, including all solver, residual, alignment, and network costs?
- 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.
- How stable are the main comparisons across independent training and noise seeds?
3 comments
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 = 2 | 20 dB | 30 dB | 40 dB |
|---|---|---|---|
| Spatial (per-pixel) blending, PhaseFuse | 30.33 | 38.46 | 47.37 |
| Whole-image soft blending | 30.15 | 38.24 | 47.33 |
| Whole-image hard selection | 29.99 | 38.07 | 47.33 |
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 = 2 | Experts chosen by whole-image selection (images out of 200) |
|---|---|
| 20 dB | TV-TWF λ = 0.02 (129), TV-TWF λ = 0.01 (71) |
| 30 dB | TV-TWF λ = 0.002 (177), TV-TWF λ = 0.005 (23) |
| 40 dB | TV-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 = 2 | Time per image | 20 dB | 30 dB | 40 dB |
|---|---|---|---|---|
| TWF projected after every update, 200 updates | 8.8 ms | 27.24 | 37.18 | 47.16 |
| TWF projected after every update, 5,000 updates | ≈ 220 ms | 27.22 | 37.18 | 47.16 |
| Best standalone solver (TV-TWF λ = 0.005), 1,000 updates | ≈ 315 ms | 29.35 | 37.36 | 40.23 |
| PhaseFuse (all solver, residual, alignment and network costs) | 320 ms | 30.33 | 38.46 | 47.37 |
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.
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.
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:
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:
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.