{"id":"0fa40a75-593b-4ff0-b4c0-5c5a9a32e7da","arxiv_id":"2411.18189","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Untrained SIREN-based implicit neural representations, optionally initialized by prior embedding, outperform Deep Decoder and low-shot decoder baselines on simulated lensless deblurring.","lead":"This paper applies implicit neural representations to lensless image deblurring, optimizing a SIREN network without pretraining so that blurring its output with the camera's point spread function matches the captured image. It reports better PSNR and SSIM than Deep Decoder and low-shot baselines, especially when the network is first initialized on a single image from the same domain.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation uses the same forward model for synthesis and reconstruction, so the reported margin may not survive PSF mismatch or real sensor noise.","rationale":"The reader's weakest_assumption correctly identifies the main vulnerability: the synthetic measurements are generated with the same forward model used for reconstruction. I see no internal inconsistency in the method itself; the SIREN optimization with a PSF-informed loss is coherent, and the under-parameterization comparison is a plausible way to control for model capacity. The prior-embedding step is also well motivated. However, the reported superiority is only as strong as the realism of the evaluation. Because the forward model is embedded in both the data-generation and the reconstruction, the experiments cannot distinguish genuine physical regularization from overfitting to the assumed PSF. The paper's own discussion in Sec. IV-A concedes that the shift-invariant PSF is an approximation, which makes this gap central rather than incidental. Baseline tuning and the unaccounted cost of prior embedding are secondary; they affect the size of the margin but not the core possibility of the method. The proposed test—using a perturbed or real PSF—would directly probe whether the margin survives. If it does, the conditional verdict can be upgraded; if not, the claim should be restricted to the exact-model setting. Thus I keep the reader's CONDITIONAL verdict unchanged.","tokens_in":12129,"tokens_out":5445,"duration_ms":51307,"concrete_test":"Acquire or simulate a test set with the same clean images but generate measurements with a perturbed PSF (e.g., the calibrated DiffuserCam PSF convolved with a small shift-variance kernel, or a 5% kernel error) and real sensor noise, while reconstructing with the original nominal PSF. Run the identical untrained SIREN and modified Deep Decoder protocols (same iterations, same UPR) and compare PSNR/SSIM. If the INR advantage over the Deep Decoder drops below the reported margin or reverses, the central claim is conditional on exact forward-model match.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that an untrained SIREN with a PSF-informed forward loss outperforms existing untrained and low-shot methods for lensless deblurring. For that claim to be substantiated, the evaluation must test the method under conditions resembling real lensless capture. In this paper, all quantitative tests (Secs. V-A, V-C) generate the measurements by applying the same FFT-convolution forward model, Eq. (14), with the same known DiffuserCam PSF and small Gaussian noise that is used inside the reconstruction loop. The generator and the solver therefore share the exact operator. This can only demonstrate performance under perfect model match. The authors themselves note that the shift-invariant PSF assumption is physically inexact and error-prone (Sec. IV-A, citing [29]), and real DiffuserCam measurements include shift variance, calibration error, and sensor noise not captured by i.i.d. Gaussian noise. Under model mismatch, an untrained network can fit artifacts of the assumed forward model, and the reported PSNR/SSIM margin over Deep Decoder and ADMM may shrink or reverse. This is an external-validity gap, not an internal inconsistency, but it is the load-bearing link between the experiments and the headline claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an untrained implicit neural representation (SIREN) for lensless image deblurring. Reconstruction is performed by optimizing the INR to minimize the discrepancy between the PSF-convolved network output and the measured lensless image, with FFT-based convolution as the forward model. The method is extended with a prior-embedding step: the INR is first fit to a single domain-specific image and then used as the initialization for the physics-informed optimization. The authors report PSNR/SSIM improvements over Deep Decoder, ADMM, Learned-ADMM-UNet, and a 10-shot domain-restricted method on synthetic lensless images generated from the DiffuserCam PSF, and introduce the Under-Parameterization Ratio (UPR) to compare network sizes in the untrained regime.","tokens_in":12335,"tokens_out":4777,"duration_ms":43918,"significance":"If the reported results were validated under realistic lensless imaging conditions, the method would be a useful data-efficient alternative to GAN-based reconstructions, with the potential to adapt to new PSFs without retraining. The UPR metric is a simple and potentially useful tool for comparing under-parameterized network families. The paper includes thorough comparisons against several baselines and clear visualizations. However, the current evaluation is entirely based on synthetic measurements generated with the same forward model that is used inside the reconstruction loss, which severely limits the external validity of the central claim of superiority. The prior-embedded comparison is also confined to a single narrow domain, making the claimed generalization to the low-data regime insufficiently supported.","major_comments":[{"comment":"All quantitative evaluations are performed on synthetic lensless images created by convolving clean images with the same DiffuserCam PSF k and adding small Gaussian noise, which is exactly the forward model used in the reconstruction loss inside the optimization loop. The authors themselves note in Sec. IV-A (citing [29]) that the shift-invariant PSF assumption is physically inexact and error-prone. Under model mismatch, such as real sensor noise, PSF calibration error, or shift variance, an untrained network can fit artifacts of the assumed forward model, and the reported margin over Deep Decoder and ADMM may shrink or reverse. To substantiate the claim of 'superior performance by a significant margin' for practical lensless imaging, the paper needs at least one experiment with real DiffuserCam measurements or with a deliberately mismatched PSF (e.g., synthesizing with a different PSF than the one used in the loss).","section":"Sec. V-A, V-C; Eq. (14)"},{"comment":"The prior-embedded experiment uses a single microorganism image as the prior and a second microorganism image as the test target, both from the same narrow domain and dataset. This setup does not demonstrate that prior embedding provides a general advantage in the low-data regime; the improvement may largely reflect that the prior and target images are nearly identically distributed. The claim of 'bridging the gap between no-data and high-data regimes' (Abstract, Sec. VI) requires evaluation across multiple domains or at least on a more diverse test set, together with a sensitivity analysis with respect to the choice of the prior image (e.g., using an unrelated-domain prior or a degraded prior).","section":"Sec. V-C; Fig. 6"},{"comment":"The reported quantitative results are single-seed and given for fixed iteration counts, but the paper provides no analysis of sensitivity to the random initialization of the INR, the learning rate, the optimizer choice, or the Gaussian noise level in the forward model. Since the method is untrained and the entire reconstruction depends on the optimization trajectory, the absence of variance or statistical significance measures makes the claim of a 'significant margin' difficult to assess. The authors should report mean and standard deviation over multiple random seeds, or at least show that the ranking of methods is stable across hyperparameter choices.","section":"Sec. V-A; Table I and Fig. 3"}],"minor_comments":[{"comment":"The notation ω(i)0 for the frequency scaling parameter of each SIREN layer is nonstandard; in the original SIREN paper, a single ω0 is used only in the first layer and subsequent layers use ω0=1. The authors should clarify whether they use a per-layer frequency scaling and how it affects initialization.","section":"Sec. IV-B, Eq. (17)"},{"comment":"The loss function is described inconsistently: Algorithm 1 says 'L1 loss function', the text in Sec. IV-C says 'mean-squared error', Eq. (10) writes an L2 norm, and Fig. 1 says 'LMSE'. The exact loss used in all experiments should be stated unambiguously.","section":"Algorithm 1 and Sec. IV-C"},{"comment":"The objective for prior embedding is written with malformed triple norms, and the subsequent equation 'θ∗ = arg min θ ||Mθ−y;Mprior||2 2' is not a well-formed optimization problem; the forward model k∗ and the initialization from Mprior should be stated explicitly.","section":"Sec. IV-D, Eq. (18)"},{"comment":"The claimed trend 'increasing SSIM with decreasing UPR' is not strictly monotonic in the reported numbers (e.g., MDD at 10k iterations: UPR 1.47 gives 0.76, UPR 3.46 gives 0.73). The discussion should acknowledge this and avoid overstating the monotonicity.","section":"Table I"},{"comment":"The 'microorganism preprocessed images dataset [33]' is only cited as a Kaggle dataset; the exact dataset name, version, and any preprocessing steps should be provided for reproducibility.","section":"Sec. V-C"},{"comment":"The claim of being 'the first to leverage INRs for lensless image reconstruction' should be supported by a more exhaustive related-work search, since there may be concurrent or earlier works using implicit neural representations for lensless or coded-aperture imaging.","section":"Sec. I and II"}],"recommendation":"major_revision","confidential_remarks":"The central technical idea is simple and plausible, and the paper is clearly written overall. The main obstacle is the evaluation methodology: every experiment uses the same forward model for both synthesizing the test measurements and reconstructing from them, so the reported margins are entirely self-consistent by construction. Because the authors themselves acknowledge (Sec. IV-A) that the shift-invariant PSF model is physically inexact, the lack of any model-mismatch or real-data experiment is a load-bearing gap. I would recommend requiring at least one experiment with a mismatched PSF or real DiffuserCam measurements, and a multi-seed variance analysis, before considering the paper for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this is a coherent paper that combines SIREN-style implicit neural representations with a known-PSF forward loss for lensless deblurring, and adds a prior-embedding trick to speed convergence. The specific combination appears new relative to the cited untrained lensless work, and the UPR analysis is a simple but useful way to compare under-parameterized networks. The paper is worth a serious look, but the empirical case for 'significant margin' is not yet made: every quantitative test uses the same FFT-convolution forward model to synthesize the measurements and to compute the reconstruction loss, with the same PSF and the same small Gaussian noise. That's a perfect model-match setup, and the authors themselves flag in Sec. IV-A that the shift-invariant PSF assumption is physically inexact and error-prone, citing [29]. So the claimed superiority over Deep Decoder and ADMM may not survive PSF mismatch, calibration error, or real sensor noise. This is an external-validity gap, not an internal inconsistency, but it's load-bearing for the headline claim.\n\nWhat the paper does well: the optimization loop is simple and clearly described; the UPR metric is a nice practical tool; the prior-embedding results, even if synthetic, show a real convergence benefit; and the related work is mostly well chosen. The comparisons to the authors' own prior methods [16], [18] are legitimate baselines, not circular inputs.\n\nSoft spots besides the forward-model issue: no code or detailed hyperparameters are released, baseline tuning is not demonstrated (so the comparison may be unfair), and the 'significant margin' language in the abstract oversells what is shown. I'd also want to see how the method behaves when the test PSF differs from the one used in the loss, or better, on real DiffuserCam measurements.\n\nWho is this for? People working on untrained reconstruction for lensless or computational imaging, and anyone interested in INR-based inverse problems in low-data settings. It deserves a serious referee, but the revision needs real or held-out PSF experiments, careful baseline tuning, and a more measured claims section.\n\nMy recommendation: send it to peer review, but with the expectation of significant revision on the evaluation. I'd not cite it myself until the external-validity question is addressed.","headline":"New combination of INR and known-PSF untrained optimization for lensless deblurring, but the synthetic same-model evaluation leaves the claimed margin unproven.","tokens_in":12888,"tokens_out":2431,"would_cite":false,"duration_ms":20725,"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":"An untrained SIREN with a PSF-informed loss outperforms low-shot lensless deblurring methods.","keywords":["lensless imaging","implicit neural representations","untrained neural networks","image deblurring","prior embedding","low-data regime","point spread function","under-parameterization"],"falsifier":"Apply the same untrained SIREN to real lensless captures from a DiffuserCam-style camera with a calibrated PSF and compare against a sharp reference; if the margin over the Deep Decoder shrinks or reverses under PSF mismatch or real sensor noise, the claimed superiority depends on the exact forward model.","tokens_in":11902,"feed_emoji":"📷","tokens_out":5363,"duration_ms":45981,"temperature":0.7,"pith_summary":"The paper tries to establish that implicit neural representations—coordinate networks that map pixel positions to color—can solve lensless image deblurring without any dataset training. Its central claim is that an untrained SIREN, optimized only against a physics-informed loss that convolves the network output with the known point-spread function, reconstructs lensless images with higher PSNR and SSIM than established untrained and low-shot alternatives. A second claim is that warm-starting the network by fitting it to a single image from the same domain, called prior embedding, speeds convergence and improves quality beyond a 10-shot domain-restricted method. If these claims hold, lensless cameras could be computationally deblurred with a compact network and a known PSF, without collecting paired training data or retraining for each camera.","feed_headline":"No training data: one SIREN network deblurs lensless images","feed_subtitle":"A known PSF in the loss lets an untrained coordinate network beat low-shot baselines.","key_machinery":"The central object is a SIREN, an MLP with sinusoidal activations that maps continuous coordinates $(u,v)$ to RGB values, kept under-parameterized relative to the $256 \\times 256 \\times 3$ image space. The mechanism that carries the argument is the physics-informed forward loss: at every optimization step the network's intermediate image is convolved with the known DiffuserCam PSF using FFT convolution, small Gaussian noise is added, and the error against the observed lensless image is backpropagated. Prior embedding is the second mechanism: before inverse optimization, the network is trained to fit one image from the target domain, producing a weight initialization that accelerates the subsequent PSF-informed reconstruction. The Under-Parameterization Ratio, defined as image-space dimensionality divided by network parameter count, supplies the comparison metric across architectures.","core_discovery":"On its own terms, the paper's discovery is that the inductive bias of a sinusoidal coordinate network is enough to turn a lensless measurement back into a clean image when the PSF is known: the network's output is convolved with the PSF in the Fourier domain and compared to the measurement, and that loss alone drives reconstruction. The authors report that this untrained SIREN outperforms the modified Deep Decoder at matched parameter counts, and that embedding a single domain image into the network's weights before optimization outperforms a 10-shot decoder baseline. They also introduce the Under-Parameterization Ratio to quantify how far below image dimensionality a network operates, and show that reconstructions improve as UPR decreases toward 1.","pith_inferences":["The paper leaves untested the natural extension that a converged coordinate network can be sampled at arbitrary resolutions, since the INR is a continuous function rather than a pixel grid.","The monotone UPR-SSIM trend suggests a practical design rule: choose network width by sweeping UPR on synthetic examples before full optimization for a new PSF.","The evaluation never exposes the method to PSF mismatch or real sensor noise, so a testable next step is to add a calibration or mismatch-robust term before claiming real-world applicability."],"forward_implications":["A roughly 132k-parameter SIREN can reconstruct 256x256 lensless images without paired training data, outperforming the untrained Deep Decoder at similar parameter counts.","Prior embedding from a single domain image gives faster convergence and better PSNR and SSIM than the 10-shot domain-restricted decoder baseline.","The Under-Parameterization Ratio provides a transferable way to compare untrained networks across architectures at matched under-parameterization.","Because the PSF is inside the loss, the same optimization procedure applies to any calibrated lensless camera by swapping the kernel, with no retraining."],"supporting_citations":[{"why":"Supplies the DiffuserCam point spread function and dataset used to synthesize the lensless measurements and as the imaging system being modeled.","marker":"[1]"},{"why":"Defines the Deep Decoder architecture, the under-parameterized non-convolutional network used as the primary untrained baseline.","marker":"[14]"},{"why":"Adapts Deep Decoder to lensless deblurring and provides the direct untrained comparison at matched parameter counts.","marker":"[16]"},{"why":"Provides the 10-shot domain-restricted low-shot method against which prior-embedded reconstruction is benchmarked.","marker":"[18]"},{"why":"Documents the non-idealities of shift-invariant PSF assumptions, the main threat to the forward-model premise.","marker":"[29]"},{"why":"Supplies the SIREN architecture with sinusoidal activations that is the implicit neural representation being optimized.","marker":"[30]"}],"fun_headline_variants":["SIREN alone deblurs lensless images with no training","Untrained SIREN beats low-shot baselines in lensless deblur","One network, zero training: lensless deblurring via SIREN","No data, no GAN: SIREN deblurs lensless images","Zero-shot lensless deblurring with a SIREN network"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The evaluation assumes the FFT-convolution forward model with the known DiffuserCam PSF and small Gaussian noise exactly generates the test measurements, because those measurements were synthesized by that same model.","fun_headline_variants_meta":{"raw":{"variants":["SIREN alone deblurs lensless images with no training","Untrained SIREN beats low-shot baselines in lensless deblur","One network, zero training: lensless deblurring via SIREN","No data, no GAN: SIREN deblurs lensless images","Zero-shot lensless deblurring with a SIREN network"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000234,"raw_usage":{"total_tokens":1468,"prompt_tokens":889,"completion_tokens":579,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":478}},"tokens_in":505,"tokens_out":579,"duration_ms":4816,"temperature":1.0,"reasoning_tokens":478,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:26:02.607502+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Apply the same untrained SIREN to real lensless captures from a DiffuserCam-style camera with a calibrated PSF and compare against a sharp reference; if the margin over the Deep Decoder shrinks or reverses under PSF mismatch or real sensor noise, the claimed superiority depends on the exact forward model.","supporting_citations":[{"cited_title":"Diffusercam: lensless single-exposure 3d imag- ing,","cited_arxiv_id":null,"evidence_quote":"Supplies the DiffuserCam point spread function and dataset used to synthesize the lensless measurements and as the imaging system being modeled."},{"cited_title":"Deep decoder: Concise image repre- sentations from untrained non-convolutional networks,","cited_arxiv_id":null,"evidence_quote":"Defines the Deep Decoder architecture, the under-parameterized non-convolutional network used as the primary untrained baseline."},{"cited_title":"Physics-informed deep deblurring: Over-parameterized vs. under-parameterized,","cited_arxiv_id":null,"evidence_quote":"Adapts Deep Decoder to lensless deblurring and provides the direct untrained comparison at matched parameter counts."},{"cited_title":"Reconstruct- ing synthetic lensless images in the low-data regime.,","cited_arxiv_id":null,"evidence_quote":"Provides the 10-shot domain-restricted low-shot method against which prior-embedded reconstruction is benchmarked."},{"cited_title":"Robust reconstruction with deep learning to handle model mismatch in lensless imaging,","cited_arxiv_id":null,"evidence_quote":"Documents the non-idealities of shift-invariant PSF assumptions, the main threat to the forward-model premise."},{"cited_title":"Implicit neural representations with periodic activation functions,","cited_arxiv_id":null,"evidence_quote":"Supplies the SIREN architecture with sinusoidal activations that is the implicit neural representation being optimized."}],"review_version":1}