{"id":"f0111086-4e48-4f30-b061-37b2ad84e5b7","arxiv_id":"2608.06574","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A training-free implicit neural representation fit with an aperture-aware maximum-likelihood loss denoises dynamic speckle video without clean references or retraining.","lead":"Spatiotemporal neural networks can remove speckle noise from dynamic coherent imagery by fitting a physics-based likelihood, without training data. The method adapts to any aperture shape and stops itself before overfitting, outperforming supervised and classic denoisers on simulated and lab holography videos.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The temporal-independence assumption in Eq. (4) is violated for translating targets, and the simulated validation generates independent speckle per frame; the dynamic-scene claim is therefore untested under physical speckle motion.","rationale":"The reader's weakest assumption points at the fully developed speckle model; I agree broadly but isolate temporal independence as the load-bearing sub-assumption. The derivation of Eq. (4) is correct under i.i.d. speckle, and the matrix-free implementation is checked against explicit covariance evaluation in Supplement 1, so the internal logic is sound. The problem is that the simulated experiments instantiate the model's own independence assumption, while the motivating application, dynamic translations of rough objects, produces translating, temporally correlated speckle. This is not a disagreement with consensus; it is a correctness risk about whether the headline comparisons transfer to the physical regime. The lab experiments partly mitigate the concern because they involve actual translating targets and the method still denoises visibly, which is why I do not move the verdict to REJECT. But the concern is concrete and testable, and it is exactly the kind of condition that should gate acceptance. The reader's verdict is already CONDITIONAL; my read adds a specific condition, a fixed-reflectivity translating-target simulation or quantitative lab validation, rather than changing the verdict.","tokens_in":17647,"tokens_out":15014,"duration_ms":148468,"concrete_test":"Replace the independent per-frame g_t in Supplement 4 with a fixed complex reflectivity: draw h ~ CN(0,I) once, set the object-plane field at frame t to a shifted copy s_t = sqrt(r_t) * h(x - d_t) (circular shift for the translating sequences), and generate y_t = A s_t + n_t for the large, small, and Golay apertures. Run the proposed method with the same hyperparameters and blind holdout on the correlated-speckle data, and compare per-sequence PSNR/SSIM/ST-RRED against Table 2 and against FastDVDNet-A. If the advantage over baselines collapses or falls within sequence-to-sequence spread, the central dynamic-scene claim is not supported; if the margins persist across all seven DERF-HD sequences, the temporal-independence assumption is not the bottleneck.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2A derives Eq. (4) under 'independent speckle and additive-noise realizations over time.' For a rigidly translating rough object, the complex reflectivity is a fixed random field; each frame's object-plane field is a shifted copy of the same field, so the image-plane speckle translates with the object and is strongly correlated across frames. Eq. (4) is then a product of per-frame marginals, a composite likelihood rather than the joint maximum likelihood, and the temporal samples do not provide the independent averaging the likelihood assumes. The simulated data (Supplement 4) draw a fresh g_t for each frame, exactly reproducing the model's assumption; this validates the algorithm under its own generative model but not under the moving-target physics the title and introduction target. The laboratory experiments do use translating targets and show qualitative improvements, which is genuine evidence of robustness, but they are not quantitative and cannot establish that the reported PSNR/SSIM/ST-RRED margins survive with realistic speckle motion. Section 4 acknowledges temporal speckle correlation as a limitation, yet the central dynamic-scene claim is precisely where this limitation is load-bearing: if the reported gains rely on temporal averaging of independent speckle, the key comparison in Table 2 is an in-model demonstration rather than a test of the claimed physical scenario.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a training-free framework for denoising dynamic coherent imagery by fitting a spatiotemporal implicit neural representation (INR) to noisy complex-field measurements with an aperture-aware maximum-likelihood objective. The likelihood models the measurement covariance as Sigma_t = A D(r_t) A^H + sigma^2 I, where A encodes pupil propagation, and is made practical through FFT-based covariance products, stochastic Lanczos quadrature for the log-determinant, and conjugate gradients for the quadratic form. A blind holdout criterion using two withheld frames provides early stopping without clean references. Experiments on simulated DERF-HD videos under large, small, and Golay apertures report improvements in PSNR, SSIM, and ST-RRED over classical (H-BM3D), unsupervised (DVP), and supervised (DRUNet, FastDVDNet) baselines, including aperture-retrained variants. Laboratory off-axis holography experiments with translating targets and circular and annular apertures show qualitative improvements in spatial detail and temporal consistency.","tokens_in":17878,"tokens_out":7932,"duration_ms":69814,"significance":"If the results hold, the work is significant: it couples coherent-imaging physics to INR priors, requires no external clean training data, adapts to known pupil geometries without retraining, and provides a matrix-free implementation that scales to realistic resolutions. The blind holdout stopping rule is a practical contribution, and the comparison against aperture-retrained supervised baselines is unusually thorough. However, the central dynamic-scene claim is weakened by the temporal-independence assumption in the likelihood: the quantitative simulated validation is performed under the model's own assumption of independent speckle per frame, while the physically motivated dynamic-scene regime involves temporally correlated speckle for translating rough objects.","major_comments":[{"comment":"The sequence-level likelihood in Eq. (4) is derived under independent speckle and additive-noise realizations over time, and the simulated measurements are generated with a fresh g_t for each frame, so Table 2 validates the method under exactly this independence assumption. For a rigidly translating rough object, the object-plane complex field is a single fixed random field shifted across frames, so the image-plane speckle is strongly correlated in time and Eq. (4) is a composite likelihood rather than the joint likelihood. This assumption is load-bearing for the paper's central dynamic-scene claim: the reported PSNR/SSIM/ST-RRED margins may rely on temporal averaging of independent speckle. Section 4 lists temporal speckle correlation as a limitation, but that is precisely the regime the title and introduction target. The laboratory translating-target experiments provide genuine qualitative evidence of robustness, but they do not quantify the margins under correlated speckle. Please add simulated experiments with physically correlated speckle (for example, a fixed complex reflectivity translated by a known motion between frames) and report the same metrics, or temper the dynamic-scene claim accordingly.","section":"Section 2A, Eq. (4); Supplement 4; Section 4"},{"comment":"The validation of the matrix-free approximation reports PSNR 22.4 versus 22.0 dB and SSIM 0.71 versus 0.69, but ST-RRED degrades from 13,486 to 22,945, roughly a 70% increase. Since all main results use the matrix-free implementation and Table 2 reports ST-RRED as a headline temporal-consistency metric, the approximation error in ST-RRED should be characterized at the resolution and aperture conditions used in Table 2, or the reported ST-RRED values should be interpreted with this caveat. The current explanation attributes the discrepancy to low resolution and metric sensitivity, but it does not establish that the effect is negligible at 256x256.","section":"Supplement 1, Fig. S1"}],"minor_comments":[{"comment":"The main text should state explicitly that the two withheld frames are included in the PSNR, SSIM, and ST-RRED evaluations; this disclosure currently appears only in the supplement.","section":"Section 2C and Supplement 1, Section 3"},{"comment":"The ablation study is performed on a single DERF-HD sequence (sunflower); including a second sequence would increase confidence that the conclusions about architecture and loss are not sequence-specific.","section":"Section 3C, Table 3"},{"comment":"The phrase 'training-free' should be clarified to mean no external training data or pretrained weights, since the INR is still optimized per sequence; the current wording could be misread as claiming no optimization at all.","section":"Section 1"},{"comment":"The text states that 'our method continues to outperform all baselines' for the Golay aperture, but the following sentence reports that PSNR matches FastDVDNet-A at 22.2 dB; please qualify the claim by noting that the improvement is in SSIM and ST-RRED rather than PSNR.","section":"Section 3B, Golay aperture paragraph"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for the journal and the central idea is sound. The main risk is the mismatch between the simulated validation, which assumes independent speckle per frame, and the dynamic-scene claim, which is most physically interesting when speckle is temporally correlated (e.g., for translating rough objects). The laboratory experiments partially mitigate this risk, but they are qualitative. I recommend major revision with a request for correlated-speckle simulations and for a more careful characterization of ST-RRED sensitivity to the matrix-free approximation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real advance and deserves refereeing. The method—fitting a spatiotemporal INR to a coherent MLE loss that models aperture-induced speckle covariance, with a matrix-free implementation and a blind holdout stopping rule—is new as a package, and the validation is mostly honest. But the headline numbers are produced under the model's own temporal-independence assumption, and the only physical moving-target evidence is qualitative. Don't take Table 2 as proof that the gains survive for speckle that translates with the scene.\n\nWhat's good: the likelihood is derived from standard Goodman speckle physics, not reverse-engineered from the results. The matrix-free trick (FFT covariance + SLQ + CG) is legit and validated against explicit evaluation (PSNR 22.4 vs 22.0; SSIM 0.71 vs 0.69). The ablations show that both the tuned WIRE backbone and the MLE loss matter: replacing MLE with intensity MSE drops PSNR from 27.9 to 25.7, and using the original WIRE hyperparameters drops it to 23.7. The baselines are strong, including aperture-retrained supervised models, and DVP is given oracle early stopping. The lab experiments with a real holography setup, including a pupil change to an annular aperture, are convincing evidence that the method is not purely a simulation artifact.\n\nSoft spots. The big one: Eq. (4) assumes independent speckle and noise realizations across time. For a rigidly translating rough object, the complex reflectivity is a fixed random field; the speckle pattern translates and is correlated frame-to-frame. The simulated data generation draws a fresh g_t per frame, so Table 2 validates the algorithm under exactly the assumption of the likelihood. That's a clean in-model demonstration, but it does not test the physical moving-target scenario. The lab experiments use translating targets and do show qualitative improvement, which is reassuring, but there are no quantitative metrics. The authors acknowledge temporal speckle correlation in Section 4, but it sits right on the main claim. I'd ask them to add a simulation where the speckle field is a single fixed random draw shifted per frame (with and without additive noise) and see how much the margin shrinks. That would either shore up the claim or bound it honestly.\n\nMinor issues: the noise variance sigma for the lab likelihood is never stated, which makes the lab results hard to reproduce. Table 2 reports averages without error bars across the seven DERF-HD sequences. No code or data is released (the data availability note says 'not publicly available at this time'), which for a methods-heavy paper is a real cost. The early-stopping patience is set separately for simulated (150) and lab (750) data; there is no sensitivity analysis. All fixable.\n\nWho's it for: computational imaging researchers working on SAR, digital holography, or OCT. I'd send it to a strong referee. It's not a desk reject. The core physics is right; the soft spots are about verification breadth and documentation, not a wrong derivation.","headline":"A genuine new combination—INR fit with an aperture-aware coherent MLE loss and a matrix-free implementation—that deserves serious refereeing, but the headline numbers are in-model because the simulations assume independent speckle per frame; the only physical moving-target evidence is qualitative.","tokens_in":18474,"tokens_out":5088,"would_cite":true,"duration_ms":44921,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Fitting a spatiotemporal implicit neural representation against an aperture-aware coherent maximum-likelihood objective recovers dynamic speckle-free reflectance from noisy coherent measurements without training data, clean references…","keywords":["speckle denoising","implicit neural representation","coherent imaging","digital holography","maximum likelihood","training-free","dynamic scenes","aperture-aware"],"falsifier":"A deciding test is to simulate a sequence in which the complex speckle field is correlated across time (so $g_t$ changes smoothly rather than independently between frames) while keeping the same pupil, then run the method with blind holdout stopping and compare the stop point against the oracle-MSE minimum; if the holdout likelihood stops far from the true optimum or the recovered reflectance falls below an aperture-matched supervised baseline, the independence assumption in the likelihood is the failure point.","tokens_in":17394,"feed_emoji":"🎥","tokens_out":12063,"duration_ms":90873,"temperature":0.7,"pith_summary":"Speckle is the granular noise that limits coherent imaging, and removing it from moving scenes is hard because there is no clean frame to average toward. The paper claims that fitting a spatiotemporal implicit neural network directly to noisy complex-field measurements, using a maximum-likelihood objective that knows the pupil's effect on speckle correlations, recovers the underlying reflectance sequence without any training data, clean references, or motion estimates. Because the same likelihood also supplies a blind holdout stopping rule, the whole pipeline runs on one measurement sequence and adapts to new pupil geometries by swapping in the pupil mask. On simulated videos the recovered sequences score higher on PSNR, SSIM, and temporal-consistency metrics than classical, unsupervised, and supervised baselines, and laboratory digital holography data confirms the improvement on real measurements.","feed_headline":"Physics-based loss removes speckle from moving coherent video","feed_subtitle":"No clean frames, motion estimates, or retraining: a per-sequence network fit removes speckle from holographic video.","key_machinery":"The load-bearing object is the aperture-aware coherent maximum-likelihood loss of Eq. (4), whose per-frame covariance $\\Sigma_t = A D(r_t) A^H + \\sigma^2 I$ encodes how the pupil $p$ broadens the point-spread function and correlates speckle across pixels. A spatiotemporal WIRE network (a wavelet implicit neural representation, a multilayer perceptron with complex Gabor activations) serves as the reflectance prior, mapping coordinates $(u,v,t)$ to nonnegative reflectance, with a tuned low oscillation frequency that favors low-frequency scene structure early in fitting. The loss is evaluated matrix-free: FFTs apply the covariance, stochastic Lanczos quadrature estimates the log-determinant, and conjugate gradients solve the inverse quadratic form, giving per-frame cost $O((SK+G)N\\log N)$. A blind holdout rule stops optimization at the minimum of the held-out likelihood.","core_discovery":"The central discovery is that the covariance structure of fully developed speckle, written as $\\Sigma_t = A D(r_t) A^H + \\sigma^2 I$ with $A = F^H D(p) F$ the coherent propagation operator through pupil $p$, turns denoising into a tractable maximum-likelihood problem when combined with an implicit neural representation. The paper argues that optimizing the network parameters $\\theta$ against this likelihood, instead of a generic pixelwise loss, is what lets the reconstruction separate shared scene reflectance from frame-specific speckle, and that the same likelihood evaluated on two held-out frames provides a reference-free early-stopping signal. The method is training-free in the sense that each sequence is denoised by a fresh optimization, and aperture changes are handled by the pupil mask in the forward model rather than by retraining.","pith_inferences":["Because the holdout MLE is claimed to track the unobservable oracle MSE, the same stopping rule could plausibly transfer to other inverse problems with known forward covariances, such as multi-look digital holography or partial-coherence tomography; the paper does not test this transfer.","A natural extension the paper leaves implicit is joint estimation of the pupil or static aberrations along with reflectance, since the likelihood already depends on $p$; if the pupil is only approximately known, the method's advantage over retrained baselines may shrink.","The reported improvement with temporal window length up to 32 frames suggests that adaptive windowing or larger INR capacity could push performance further on longer sequences, but the paper fixes window size and architecture.","The matrix-free approximations (stochastic Lanczos quadrature and conjugate gradients) introduce a small PSNR loss relative to the explicit likelihood; replacing them with cheaper deterministic estimators could reduce runtime further, though the paper does not explore that direction."],"forward_implications":["A single noisy measurement sequence, plus knowledge of the pupil, is enough to denoise dynamic coherent imagery; no clean reference frames, external datasets, motion estimates, or aperture-specific retraining are needed.","The same objective transfers to different pupil geometries—large, small, annular, and sparse Golay apertures—by changing only the pupil mask $p$, so a new sensor geometry does not require collecting matched training data.","Blind holdout early stopping makes the per-sequence optimization fully automatic, removing model selection as a practical barrier.","On simulated DERF-HD videos the method reports the best PSNR, SSIM, and ST-RRED among compared methods for large and small apertures, and on laboratory digital holography it preserves fine resolution-bar structure and temporal stability.","The matrix-free likelihood evaluation scales from explicit-covariance infeasibility at $256\\times 256$ to practical operation at $512\\times 512$, with per-frame denoising runtimes near 85 seconds at $256\\times 256$."],"supporting_citations":[{"why":"Supplies the fully developed speckle field model $r_t^{1/2}\\circ g_t$ with $g_t\\sim CN(0,I)$ that the likelihood is built on.","marker":"[1]"},{"why":"Establishes the speckle statistics and spatial-correlation behavior that motivate modeling aperture-dependent covariance rather than pixelwise noise.","marker":"[2]"},{"why":"Provides the aperture-aware coherent modeling result that motivates retaining the coherent transfer function in the likelihood.","marker":"[10]"},{"why":"Supplies the training-free video-prior baseline that the paper compares against and contrasts with its likelihood-based objective.","marker":"[18]"},{"why":"Supplies the WIRE architecture with complex Gabor activations used as the spatiotemporal reflectance prior.","marker":"[27]"},{"why":"Defines the BM3D denoiser used inside the homomorphic classical baseline.","marker":"[28]"},{"why":"Defines the DRUNet supervised single-frame baseline; the aperture-aware variant retrains the same architecture.","marker":"[30]"},{"why":"Defines the FastDVDNet supervised video-denoising baseline; the aperture-aware variant retrains the same architecture.","marker":"[31]"}],"fun_headline_variants":["Training-free neural fit removes speckle from video","No clean frames needed: speckle removed from motion video","Per-sequence network fit beats speckle without retraining","Aperture-aware likelihood denoises dynamic scenes","Blind holdout early stop: speckle-free video fit"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the scene is a matte, rough reflector whose speckle noise is independent from pixel to pixel and from frame to frame, and that the optical path adds no time-varying phase errors; if those conditions fail, the covariance model the loss is built on no longer describes the measurements.","fun_headline_variants_meta":{"raw":{"variants":["Training-free neural fit removes speckle from video","No clean frames needed: speckle removed from motion video","Per-sequence network fit beats speckle without retraining","Aperture-aware likelihood denoises dynamic scenes","Blind holdout early stop: speckle-free video fit"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000546,"raw_usage":{"total_tokens":2565,"prompt_tokens":857,"completion_tokens":1708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":1631}},"tokens_in":473,"tokens_out":1708,"duration_ms":10152,"temperature":1.0,"reasoning_tokens":1631,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T14:31:55.844981+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A deciding test is to simulate a sequence in which the complex speckle field is correlated across time (so $g_t$ changes smoothly rather than independently between frames) while keeping the same pupil, then run the method with blind holdout stopping and compare the stop point against the oracle-MSE minimum; if the holdout likelihood stops far from the true optimum or the recovered reflectance falls below an aperture-matched supervised baseline, the independence assumption in the likelihood is the failure point.","supporting_citations":[{"cited_title":"Multilook Coherent Imaging: Theoretical Guarantees and Algorithms","cited_arxiv_id":"2505.23594","evidence_quote":"Supplies the fully developed speckle field model $r_t^{1/2}\\circ g_t$ with $g_t\\sim CN(0,I)$ that the likelihood is built on."},{"cited_title":"Fast estimation oftr(f(a)) via stochastic lanczos quadrature,","cited_arxiv_id":null,"evidence_quote":"Establishes the speckle statistics and spatial-correlation behavior that motivate modeling aperture-dependent covariance rather than pixelwise noise."},{"cited_title":"Plug-and-play image restoration with deep denoiser prior,","cited_arxiv_id":null,"evidence_quote":"Provides the aperture-aware coherent modeling result that motivates retaining the coherent transfer function in the likelihood."},{"cited_title":"A golay metalens for long-range, large aperture, thermal imaging via sparse aperture computational imaging,","cited_arxiv_id":null,"evidence_quote":"Supplies the training-free video-prior baseline that the paper compares against and contrasts with its likelihood-based objective."}],"review_version":2}