{"id":"7ac5e109-adbd-4f1d-a456-e312ac2dc0d4","arxiv_id":"2508.13907","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"NeuSee jointly learns a phase mask and a restoration network that, in simulation, lets a camera see a scene while withstanding laser irradiance up to 10^6 times the sensor saturation level across the visible spectrum.","lead":"NeuSee is a computational imaging system that combines a learned optical phase mask with an AI restoration network to protect cameras from laser dazzle and damage. Its main result, shown in simulation only, is that a camera can keep imaging a scene while withstanding laser light up to one million times the sensor saturation level across the visible spectrum.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 10^6× saturation suppression claim is never quantified: no absolute LSR is reported, the only relative comparison is internally inconsistent (5× vs 100×), and the DOE loss is formulated over integrated energy, not peak irradiance.","rationale":"The reader's weakest assumption is sim-to-real fidelity of the physics simulator. I agree that is unresolved and important, but the more immediate load-bearing gap is that the central number—10^6× I_sat suppression—is never directly quantified in the simulation either. The paper defines LSR in Section 3.4 and trains a DOE loss in Section 4, but Section 6 reports only a relative comparison to a half-ring mask, and the relative factor is stated as 5× in the text and 100× in the figure caption. The 10.1% L1 improvement measures restored image quality, not the physical suppression of peak irradiance, so it cannot validate the sensor-protection claim. The DOE objective in Eq. (13) is an integrated energy ratio rather than a peak-ratio objective, so even the training signal may not enforce the 10^-6 peak requirement. This is a precise, testable issue: reporting absolute LSR values or computing the post-DOE peak ratio would settle it purely in simulation, before any hardware concerns. I therefore keep the reader's CONDITIONAL verdict unchanged, but the condition should explicitly require an absolute peak-suppression evaluation in addition to hardware validation. My concern differs from the reader's sim-to-real emphasis, hence partial agreement.","tokens_in":14910,"tokens_out":15131,"duration_ms":161909,"concrete_test":"Run the trained NeuSee DOE at αl=10^6 for laser wavelengths 400–700 nm (10 nm steps), compute the post-DOE peak sensor irradiance from Eq. (7b) with the converged LSR at each wavelength, and verify max(Il,peak)/Isat ≤ 1. Also report the absolute LSR of the half-ring mask at the same wavelengths to resolve whether the relative gain is 5× or 100×. If the post-DOE peak exceeds Isat at any wavelength, the abstract's 10^6-suppression claim is false even in simulation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"NeuSee's central quantitative claim—suppressing peak laser irradiance up to 10^6× I_sat—is not directly supported anywhere in the paper. From Eqs. (6)–(7), post-DOE peak irradiance is αl·LSR·Isat, so the claim requires LSR ≤ 10^-6 across 400–700 nm. Section 6 reports no absolute LSR; it only gives a relative comparison to the half-ring mask, and that comparison is contradictory: the text says 5× stronger suppression while the Fig. 6 caption says 100×. The 10.1% L1 improvement is a restoration-quality metric, not a sensor-protection metric, so it cannot substitute for an absolute peak-suppression measurement. Additionally, the DOE loss in Eq. (13), LDOE(LSR)=Σ Il(λ)/Il0(λ), is written as an integrated on-sensor energy ratio, not a peak-ratio objective; minimizing it can be satisfied by spreading energy across pixels even if the peak suppression is far below 10^-6. The conclusion acknowledges this is a simulation result, but the abstract and introduction present the 10^6 suppression as an achieved fact without that qualifier. Until an absolute LSR is reported and the 5×/100× discrepancy is resolved, the headline claim is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces NeuSee, an end-to-end learned computational imaging framework that jointly optimizes a diffractive optical element (DOE) represented by a UNet and a frequency-space Mamba-GAN image restoration network. The system is trained on 100K simulated RGB images converted to 31-band hyperspectral radiance, with a physics-based forward model that adds laser dazzle, lens flare, sensor noise, and saturation. The authors claim that NeuSee suppresses peak laser irradiance up to 10^6 times the sensor saturation threshold, works across the full visible spectrum (400–700 nm), and improves restored image quality by 10.1% over a half-ring DOE baseline.","tokens_in":15337,"tokens_out":3282,"duration_ms":33469,"significance":"If the central claims are verified, the work would be a meaningful advance in computational imaging for laser protection, combining learned diffractive optics with learned restoration and addressing a realistic threat model (dynamic laser wavelength, intensity, position, and ambient conditions). The paper ships a detailed simulation pipeline, a sizable 100K-image training regimen, and a two-stage training strategy to balance conflicting DOE and restoration objectives. Strengths include the explicit treatment of hyperspectral propagation, sensor noise statistics, and the use of adversarial loss for restoring saturated regions. However, the headline quantitative claim — suppression of peak laser irradiance up to 10^6× I_sat — is not directly evidenced anywhere in the manuscript, and the evaluation is entirely simulation-based with no hardware verification. The significance is therefore conditional on the authors supplying the missing absolute suppression metric and reconciling internal inconsistencies.","major_comments":[{"comment":"The central claim of suppressing peak laser irradiance up to 10^6× I_sat is not supported by any reported absolute LSR value. From Eq. (7b), Il,peak = αl·LSR·Isat, so for αl = 10^6 the required LSR is at most 10^-6. Although Section 6 states that NeuSee achieves '5× stronger suppression' (text) and '100 times stronger suppression' (Fig. 6 caption) than the half-ring mask, no absolute LSR value or its wavelength dependence is reported. Moreover, the DOE objective in Eq. (13), LDOE(LSR)=Σ Il(λ)/Il0(λ), is an integrated on-sensor energy ratio, not a peak-irradiance ratio; minimizing this sum can be satisfied by redistributing energy over many pixels even when the peak LSR remains far above 10^-6. The paper must report the actual peak LSR as a function of wavelength, clarify whether the 10^6 figure refers to the training range [0, 2e6] in Section 5 or to achieved suppression, and reconcile the 5×/100× discrepancy.","section":"Section 6, Eq. (7b) and Eq. (13)"},{"comment":"The paper's claims are presented in the abstract and introduction as achieved sensor-protection capabilities, but the entire evaluation is simulated. Section 3.4 states that simulation parameter values 'match the experiment', yet no experimental setup, fabricated DOE, measured PSF, or real camera image appears in the paper. The conclusion appropriately says 'In simulation', but the abstract and introduction do not carry this qualifier. Because the learned DOE and restoration network were trained entirely in simulation, the transfer of these results to real hardware is an untested assumption. The authors should either add hardware validation or consistently qualify all headline claims as simulation-based.","section":"Sections 3.4 and 7"},{"comment":"The claim that NeuSee 'outperforms other learned DOEs' is supported only by a comparison to a single half-ring mask trained with a heuristic method [30]. The paper does not compare against other learned DOEs, such as those in Refs. [31] and [33], nor against a restoration-only baseline without any DOE. The quantitative 10.1% L1 improvement is reported as a single average over 7K test images without error bars, per-condition breakdown, or statistical significance testing. Given the strong comparative wording in the contributions list, the evaluation should include at least one additional learned-DOE baseline and report variability over test conditions.","section":"Section 6, 'outperforms other learned DOEs'"}],"minor_comments":[{"comment":"The phrase 'suppress the peak laser irradiance as high as 10^6 times the sensor saturation threshold' is ambiguous: it could mean the system handles incident lasers up to 10^6× I_sat or that it achieves an LSR of 10^-6. Please rephrase to state the intended meaning explicitly.","section":"Abstract"},{"comment":"The laser is modeled as a Dirac delta function with no spatial extent; real laser beams have finite spot size and angular divergence. This simplification should be stated as a limitation, and its effect on the reported suppression ratios should be discussed.","section":"Section 3.3, Eq. (5b)"},{"comment":"The description 'Laser strengths αl are randomly sampled from 100K predetermined values, which are uniformly distributed in the range [0, 2e6]' is unclear: are the same 100K values reused across training iterations, or independently re-sampled each iteration? Please clarify.","section":"Section 5, paragraph beginning 'Deep learning systems'"},{"comment":"The caption states '100 times stronger suppression' while the running text states '5× stronger suppression'. These should be reconciled, ideally with the numerical values of LSR for both systems included in the figure or table.","section":"Section 6, Fig. 6 caption"},{"comment":"There are formatting errors: Eq. (12b) contains 'DL 2 (bL)' where the square is misplaced, and Eq. (14) has an unbalanced parenthesis in the FFT objective ('|F(bL)−F (ˆbL)|'). Correct these for clarity.","section":"Section 4, Eq. (12b) and Eq. (14)"},{"comment":"Reference [92] is a GitHub URL without version, commit hash, or date of access; the 'Scaled Fresnel method' should be cited to a peer-reviewed publication with a precise algorithm description.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty claim of being the 'first full-spectrum' learned DOE for laser protection should be carefully checked against Refs. [31] and [33]. The absence of any absolute LSR measurement, combined with the 5×/100× inconsistency and the purely simulation-based evaluation, makes the headline result unverifiable in its current form; this is fixable with additional reporting and analysis, which is why I am not recommending rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know first: this is a well-built simulation study of a learned DOE plus restoration network for laser protection, and the combination is genuinely new. But the banner claim—suppression of 10^6× saturation across 400–700 nm—is not actually demonstrated in the paper. The stress-test note is correct: no absolute LSR is reported, the only relative number conflicts (5× in text, 100× in Fig. 6 caption), and the DOE loss in Eq. 13 minimizes an integrated energy ratio, not peak irradiance. That doesn't sink the paper, but it means the abstract overstates what is shown.\n\nWhat's new and good: the joint training of a neural phase mask with a frequency-space Mamba-GAN restorer, the two-stage strategy to separate laser suppression from image recovery, and the 100K-image physics-based synthesis pipeline with realistic noise and lens flare. The simulation physics is described in unusual detail—wavelength-dependent refractive index, scaled Fresnel propagation, full well capacity, read noise, digitization. For a purely computational study, the quantitative gain over the half-ring mask (10.1% L1) is a legitimate result, though only against one baseline.\n\nSoft spots, in order of severity. First, the missing hardware: Section 3.4 says parameters 'match the experiment,' but there is no fabricated DOE, no measured PSF, no real camera image. Sim-to-real transfer for a learned phase mask is exactly where these systems usually fail, so the central practical claim is unvalidated. Second, the 10^6 claim: given the loss formulation, the network may spread laser energy over many pixels to lower the integrated ratio without achieving the peak suppression implied by 10^6×. The paper never shows the actual peak-to-saturation number. Third, despite claiming to outperform other learned DOEs, it only compares with the heuristic half-ring mask; the learned baselines in [31–33] are cited but not tested. Fourth, no error bars, no code, no data. Fifth, the suppression ratio is a training objective, so reporting low LSR is a constraint-satisfaction result, not an independent discovery—worth stating but not a fatal flaw.\n\nWho this is for: computational imaging researchers working on learned optics and laser protection. It is worth a serious referee; the core idea and pipeline are credible. But acceptance should be conditional on major revision: report absolute LSR across wavelengths, fix the 5×/100× discrepancy, add a real experiment or at least a calibrated simulator validation, and run the missing baseline comparisons.\n\nMy recommendation: send it out, with a clear request for those revisions.","headline":"A credible simulation study of jointly learned laser-protection optics and restoration whose headline 10^6× suppression claim is unverified and internally inconsistent; worth peer review with major revisions.","tokens_in":15761,"tokens_out":2138,"would_cite":false,"duration_ms":21161,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"NeuSee jointly learns a pupil-plane phase mask and a Mamba-GAN restorer, letting cameras image through laser irradiance up to one million times the sensor saturation threshold across the full visible spectrum.","keywords":["learned diffractive optical element","computational imaging","laser dazzle protection","sensor saturation","Mamba-GAN","image restoration","full-spectrum imaging","phase mask"],"falsifier":"Fabricate the learned DOE height map with the stated material dispersion, mount it in front of a camera whose parameters match Table 1, and measure the focal-plane point spread function and the peak irradiance from a 10 nm-bandwidth laser at $\\alpha_l = 10^6 I_{\\mathrm{sat}}$; if the measured suppression ratio or restored-image L1 diverges from the simulated values by more than the assumed noise levels, the central claim fails. A faster check is to measure the fabricated DOE's point spread function on an optical bench and compare it with the point spread function predicted by Eq. 4.","tokens_in":14697,"feed_emoji":"🛡️","tokens_out":11722,"duration_ms":105589,"temperature":0.7,"pith_summary":"The paper introduces NeuSee, a computational imaging framework that pairs a learned diffractive optical element (a pupil-plane phase mask) with a frequency-space Mamba-GAN image-restoration network, trained end-to-end on 100K simulated images. Its central claim is that a camera equipped with this mask can keep imaging a scene while a laser at up to $10^6$ times the sensor saturation threshold $I_{\\mathrm{sat}}$ strikes the sensor, across the full visible band (400-700 nm). The DOE spreads and suppresses the laser's peak irradiance before it reaches the sensor, and the restoration network inpaints residual saturation, removes blur and noise, and recovers the scene radiance. The authors report that NeuSee outperforms a heuristically learned half-ring DOE, improving restored-image quality by 10.1% (L1) and giving roughly 100 times stronger average suppression across visible laser wavelengths.","feed_headline":"Learned mask plus AI restorer sees through million-fold laser dazzle","feed_subtitle":"NeuSee trains a diffractive element and a Mamba-GAN together so cameras keep imaging under 400-700 nm laser attack.","key_machinery":"The load-bearing object is the learned DOE height map $h_{\\mathrm{DOE}}(u,v)$, generated by a UNet from pupil coordinates $(u,v)$, with phase $\\phi_{\\mathrm{DOE}}(u,v) = 2\\pi\\Delta n(\\lambda) h_{\\mathrm{DOE}}(u,v)/\\lambda$. A Scaled Fresnel propagation turns this height map into wavelength-dependent point spread functions, so the mask is optimized against the full 31-band hyperspectral scene volume rather than three RGB channels. The laser is modeled as a plane wave that lands as a delta-function-like spot at the focal plane, while the background is convolved with the coded point spread function. Training uses two stages: Stage 1 jointly optimizes the mask and the restoration generator against multiscale discriminators with laser-suppression and background-transmission losses; Stage 2 freezes the DOE and fine-tunes the 8-layer FFT-Mamba restoration network with Charbonnier and Fourier-domain reconstruction losses.","core_discovery":"NeuSee's central claim is that a single learned phase mask can simultaneously scatter an intense narrowband laser so its focal-plane peak falls below the damage threshold while ordinary scene light passes through, and that the residual saturated, blurred, noisy sensor image can be restored by a Mamba-GAN trained jointly with the mask. The paper claims suppression of peak laser irradiance up to $10^6$ times $I_{\\mathrm{sat}}$, for laser wavelengths from 400 nm to 700 nm, under dynamically varying laser wavelength, intensity, position, ambient light, and sensor noise. In the reported simulations this full-spectrum protection comes with restored image quality 10.1% better on the L1 metric than the half-ring DOE baseline restored by the same network. The paper presents this as the first learned computational-imaging framework to achieve high-fidelity sensor protection across the whole visible spectrum.","pith_inferences":["The obvious next test is fabrication: etching the learned height map and measuring the on-bench point spread function and laser suppression ratio, since the paper contains no real experiment or measured point spread function.","The framework should extend to simultaneous multi-wavelength or out-of-band lasers by retraining with additional spectral bands, because the architecture is wavelength-agnostic apart from the material dispersion model.","The same joint mask-plus-restoration recipe could be applied to other saturation sources, such as sun glare or high-dynamic-range clipping, where the mask spreads the energy and the network inpaints the clipped region.","The 100 times average suppression advantage over the half-ring mask is a simulated quantity; if the simulator's delta-function laser model overstates the focused peak, real-world suppression could be lower, so the $10^6$ figure should be read as simulation-bound until hardware validation."],"forward_implications":["If the central claim holds, a single passive pupil-plane mask can replace wavelength-specific optical limiters for visible-band laser protection, since the same DOE covers 400-700 nm.","A camera using NeuSee would keep functioning at laser irradiances up to $10^6 I_{\\mathrm{sat}}$, the regime where silicon sensors begin to risk permanent damage, so the protection is not just dazzle reduction.","Because the only latency is the restoration network's post-processing, the optical mask itself provides instantaneous, linear, broadband protection without moving parts or power.","The 10.1% quality gain over the half-ring baseline indicates that jointly learned masks encode scene information more efficiently than heuristic mask shapes, suggesting further gains from richer mask parameterizations.","Deployed on autonomous vehicles, robots, security cameras, or augmented-reality headsets, the system would preserve vision during deliberate laser attacks or accidental laser exposure rather than blinding the platform."],"supporting_citations":[{"why":"Supplies the pretrained spectral-reconstruction network that converts RGB radiance into the 31-band hyperspectral volume the full-spectrum simulation needs.","marker":"[91]"},{"why":"Defines the prior learned-DOE half-ring approach that NeuSee compares against and outperforms.","marker":"[30]"},{"why":"Introduces the half-ring point spread function family used as the heuristic baseline mask.","marker":"[28]"},{"why":"Provides the Scaled Fresnel propagation implementation used to compute pupil-to-sensor point spread functions efficiently.","marker":"[92]"},{"why":"Establishes the laser-safety calculation linking saturation to damage thresholds, justifying the $10^6 I_{\\mathrm{sat}}$ target.","marker":"[15]"},{"why":"Supplies measured damage thresholds for silicon cameras under in-band and out-of-band laser exposure.","marker":"[16]"}],"fun_headline_variants":["NeuSee's AI mask blocks million-fold laser flare, restores scene","Full-spectrum laser defense: NeuSee learns mask and AI restorer","NeuSee: joint optics-AI system sees through million-fold laser dazzle","Learned mask + Mamba-GAN lets cameras see under 1e6x laser attack"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The physics-based simulator is faithful enough to a real fabricated DOE and camera that a mask trained purely in simulation will suppress a real laser at one million times saturation and let the restoration network recover the scene.","fun_headline_variants_meta":{"raw":{"variants":["NeuSee's AI mask blocks million-fold laser flare, restores scene","Full-spectrum laser defense: NeuSee learns mask and AI restorer","NeuSee: joint optics-AI system sees through million-fold laser dazzle","Learned mask + Mamba-GAN lets cameras see under 1e6x laser attack"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00101,"raw_usage":{"total_tokens":4261,"prompt_tokens":928,"completion_tokens":3333,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":544,"completion_tokens_details":{"reasoning_tokens":3248}},"tokens_in":544,"tokens_out":3333,"duration_ms":26165,"temperature":1.0,"reasoning_tokens":3248,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:10:46.997892+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Fabricate the learned DOE height map with the stated material dispersion, mount it in front of a camera whose parameters match Table 1, and measure the focal-plane point spread function and the peak irradiance from a 10 nm-bandwidth laser at $\\alpha_l = 10^6 I_{\\mathrm{sat}}$; if the measured suppression ratio or restored-image L1 diverges from the simulated values by more than the assumed noise levels, the central claim fails. A faster check is to measure the fabricated DOE's point spread function on an optical bench and compare it with the point spread function predicted by Eq. 4.","supporting_citations":[{"cited_title":"Mst++: Multi-stage spectral-wise transformer for efficient spectral reconstruction","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained spectral-reconstruction network that converts RGB radiance into the 31-band hyperspectral volume the full-spectrum simulation needs."},{"cited_title":"Learning to See Through Dazzle","cited_arxiv_id":"2402.15919","evidence_quote":"Defines the prior learned-DOE half-ring approach that NeuSee compares against and outperforms."},{"cited_title":"Half-ring point spread functions","cited_arxiv_id":null,"evidence_quote":"Introduces the half-ring point spread function family used as the heuristic baseline mask."},{"cited_title":"com / rafael - fuente / diffractsim","cited_arxiv_id":null,"evidence_quote":"Provides the Scaled Fresnel propagation implementation used to compute pupil-to-sensor point spread functions efficiently."}],"review_version":2}