{"id":"889ec526-1c5d-4ef9-aa93-6a49fea4dcd1","arxiv_id":"1908.11502","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Unrolling ADMM with learned parameters reconstructs DiffuserCam lensless images 20x faster and with better LPIPS perceptual similarity than classical ADMM.","lead":"This paper trains unrolled optimization networks to reconstruct images from mask-based lensless cameras, using paired lensed and lensless images. It reports 20x faster reconstructions than classical ADMM and better perceptual quality, with a prototype camera tested on natural scenes.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 3x LPIPS improvement is not independent evidence of perceptual quality, because LPIPS is both the training loss and the evaluation metric; an independent perceptual evaluation is needed before the central quality claim can stand.","rationale":"The reader correctly flags the transfer from monitor training to wild scenes as a key unverified assumption, and that concern is real. However, the single most load-bearing weak point for the central claim is that the only quantitative evidence for the headline quality improvement uses LPIPS, which is also the training loss. This does not invalidate the speed claim or the value of unrolling, but it means the '3x better image quality' statement is not independently established. Both concerns are addressable with additional experiments, so the conditional verdict stands. I disagree with the reader's choice of weakest assumption insofar as the LPIPS circularity is more directly tied to the strongest quantitative claim; the transfer issue is important but acknowledged as preliminary.","tokens_in":12984,"tokens_out":6999,"duration_ms":71766,"concrete_test":"Run a two-alternative forced-choice human study (or an untrained perceptual metric such as DISTS) comparing Le-ADMM-U and converged ADMM reconstructions on the existing 1,000-image test set. If human observers do not significantly prefer Le-ADMM-U, or if the LPIPS advantage does not transfer to an independent metric, the 3x LPIPS claim should be reported as an optimization result rather than a perceptual-quality improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Sec. 6.1, Table 2) is that Le-ADMM-U achieves 3x better LPIPS than converged ADMM. But Sec. 4.5 states that training minimizes a combination of MSE and LPIPS, and Sec. 6.1 evaluates with the same LPIPS metric. The comparison therefore measures how well the network optimizes its own training objective, not independently whether reconstructions are perceptually better. This is especially consequential because Le-ADMM-U's data fidelity is much worse (22.14 vs 13.62), so the network is free to add plausible texture that lowers LPIPS but may be hallucinated. The visual examples are suggestive, but the only quantitative support for 'better perceptual image quality' is a metric that the method was explicitly trained to minimize. Without an independent perceptual evaluation, the headline quality improvement may be an artifact of metric overfitting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes unrolled ADMM networks for DiffuserCam lensless imaging, with three variants: Le-ADMM (learned penalty and sparsity parameters), Le-ADMM* (adds a learned CNN regularizer), and Le-ADMM-U (adds a jointly trained U-Net denoiser). The networks are trained on 25,000 experimentally captured lensed/lensless image pairs displayed on a monitor, then compared with classical ADMM and a purely deep U-Net on a 1,000-image held-out test set. The paper reports a 20x reconstruction speedup for Le-ADMM and Le-ADMM-U, and a 3x improvement in LPIPS for Le-ADMM-U over converged ADMM, and shows qualitative reconstructions of natural objects in the wild.","tokens_in":13204,"tokens_out":4181,"duration_ms":43571,"significance":"If the central claims hold, this is a practically valuable contribution: it demonstrates a physically informed, bounded-compute reconstruction framework for a real lensless camera, provides a sizable experimental paired dataset, and makes the source code available. The explicit comparison along a model-based-to-deep spectrum is useful, and the result that a small number of learned ADMM parameters already matches classical ADMM at 20x lower cost is credible and worth publishing. The main caveat is that the headline perceptual-quality claim rests on a metric that is also used as the training loss, and the wild generalization claim is only qualitative; both need additional evidence before the claims can stand as stated.","major_comments":[{"comment":"The headline 3x LPIPS improvement of Le-ADMM-U over converged ADMM is partly by construction. The training loss in Sec. 4.5 is a combination of MSE and LPIPS, and the evaluation in Table 2 uses the same LPIPS metric. Le-ADMM-U also has markedly worse data fidelity (22.14 vs. 13.62), so it is free to generate plausible texture that lowers LPIPS without being faithful to the measurement. An independent perceptual evaluation is needed: for example, a human observer study, a perceptual metric that was not used in training, or an evaluation of reconstructions trained with an MSE-only objective. Without such evidence, the phrase 'better perceptual image quality' should be tempered.","section":"Sec. 4.5, Sec. 6.1, Table 2"},{"comment":"The claim that the networks generalize to images in the wild is supported only by qualitative visual comparisons, with no ground-truth images or quantitative metrics for the natural scenes. Because the training data are flat monitor images at a fixed focal distance, the transfer to 3D objects with variable lighting and depth is a load-bearing premise. I would like to see a quantitative evaluation on natural scenes, for example by capturing simultaneous lensed ground truth for a small set of wild scenes or by imaging known targets with registered reference images.","section":"Sec. 6.2, Fig. 6"},{"comment":"The ADMM baseline is labeled 'converged' at 100 iterations, but no convergence criterion or residual values are reported. If the baseline is not fully converged, the reported LPIPS improvement factor could be inflated. Please report the ADMM stopping criterion, primal/dual residuals, or rename the baseline '100-iteration ADMM' to avoid overstating the comparison.","section":"Sec. 6.1, Table 2"}],"minor_comments":[{"comment":"The heading 'Learned AMMM (Le-ADMM)' appears to contain a typo; it should be 'Learned ADMM'.","section":"Sec. 4.1"},{"comment":"The MSE/LPIPS loss weighting schedule is described only qualitatively ('weight MSE more heavily during earlier epochs'); for reproducibility, the exact schedule or a pointer to the released code settings should be provided.","section":"Sec. 5"},{"comment":"The U-Net data fidelity entry is listed as '19', which is ambiguous and likely missing a decimal point; please format all numbers consistently.","section":"Table 2"},{"comment":"The subfigure labels in Fig. 5(b) appear corrupted or misformatted ('10/zero.sups', '1 2 3 4', etc.); these should be fixed for legibility.","section":"Fig. 5"}],"recommendation":"major_revision","confidential_remarks":"The paper is a good fit for the journal and the experimental dataset plus code are valuable. The main risk is that the LPIPS-based quality claim is overinterpreted given the training-objective circularity; an independent perceptual evaluation and a quantitative wild-scene test would resolve this. The ADMM convergence labeling should also be corrected."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What you should know: this paper is a genuinely useful piece of applied computational imaging. It unrolls ADMM for a DiffuserCam lensless camera, learns the parameters from 25,000 experimentally captured image pairs, and gets a 20x speedup with visibly cleaner reconstructions than converged ADMM. It also ships code and gives enough setup detail to reproduce the dataset. The main caveat is the headline \"3x better LPIPS\": LPIPS is used both as the training loss (Sec. 4.5) and the evaluation metric (Sec. 6.1), so that number is partly by construction. That doesn't sink the paper, but it means the perceptual claim is weaker than it looks.\n\nThe new and good parts: this is the first application of unrolled ADMM with learned parameters to mask-based lensless imaging on a real experimental system. The paper carefully spans the model-to-deep spectrum with Le-ADMM, Le-ADMM*, Le-ADMM-U, and a plain U-Net, and shows that the intermediate approach reaches most of the quality gain with far fewer parameters and better data efficiency (Fig. 7). The training setup using a screen and beamsplitter is practical, the alignment and calibration steps are specific, and the wild tests, while qualitative, at least show the network generalizes beyond the monitor. The MSE numbers also improve substantially (0.0622 to 0.0074 for Le-ADMM-U), which suggests the quality gain isn't purely metric gaming.\n\nSoft spots, in proportion. The LPIPS circularity is real and should be fixed before publication: either train with MSE only or evaluate with an independent perceptual metric (or human ratings). The paper's own data fidelity numbers show Le-ADMM-U deviates from the physical model (22.14 vs 13.62), so the network is adding plausible texture, and we can't tell from LPIPS alone how much is real signal versus hallucination. Minor: Table 2 has no error bars or significance tests despite a 1,000-image test set. The wild generalization section is explicitly qualitative, which is fine for a preliminary demonstration, but the claim \"generalizes to natural images\" should be hedged accordingly. The discussion section does honestly acknowledge the trade-off between image quality and data fidelity, which I appreciate.\n\nWho this is for: computational imaging researchers interested in lensless cameras or unrolled optimization. It's a solid engineering contribution, not a conceptual breakthrough, but it is careful and reproducible. I'd send it to review, with the request that the authors add an independent perceptual evaluation and error bars, and soften the LPIPS claim to what the evidence actually supports.","headline":"A solid, reproducible demonstration that unrolled ADMM makes lensless imaging 20x faster with better-looking outputs, though the headline 3x LPIPS gain is inflated because LPIPS is also the training loss.","tokens_in":13725,"tokens_out":1927,"would_cite":true,"duration_ms":19665,"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":"Unrolled ADMM networks make mask-based lensless cameras practical for interactive use, reconstructing in 75 ms with better perceptual quality than a converged solver.","keywords":["lensless imaging","DiffuserCam","unrolled optimization","ADMM","learned reconstruction","perceptual loss","computational photography","deep learning"],"falsifier":"Measure a point source with the same DiffuserCam at several distances, such as 10 cm, 50 cm, and 2 m, and compare the recorded PSFs: if the pattern changes shape with depth instead of simply translating, the shift-invariant forward model in Eq. (1) is violated. A complementary test is to collect aligned lensed/lensless pairs of natural 3D scenes and compute LPIPS and MSE for Le-ADMM-U versus converged ADMM; if the learned advantage disappears outside the monitor plane, the generalization claim is not supported.","tokens_in":12809,"feed_emoji":"📷","tokens_out":8005,"duration_ms":74861,"temperature":0.7,"pith_summary":"The paper argues that the slow, artifact-prone reconstruction step that has kept mask-based lensless cameras out of everyday use can be replaced by a bounded-compute neural network that keeps the camera's known physics inside the architecture. It unrolls the ADMM optimizer into a fixed number of trainable layers, learns the penalty parameters from experimentally captured lensed/lensless image pairs, and optionally appends a trained denoiser. On a DiffuserCam prototype the resulting networks run about 20× faster than converged ADMM while matching or beating its image quality, with the deepest variant, Le-ADMM-U, reducing the LPIPS perceptual distance roughly threefold. If this holds, lensless imagers become fast enough for interactive preview and visually competitive with lensed cameras despite having no lens.","feed_headline":"Unrolled ADMM speeds lensless imaging 20× and boosts quality","feed_subtitle":"A physics-aware network reconstructs DiffuserCam images in 75 ms instead of 1.5 s, with better perceptual similarity.","key_machinery":"The central object is the unrolled ADMM network, in which each iteration of the alternating direction method of multipliers is treated as one layer of a neural network. The update equations—soft-thresholding for sparsity, least-squares for data fidelity, non-negativity projection, and dual-variable updates—stay intact, but the per-layer penalty parameters $\\mu_1,\\mu_2,\\mu_3,\\tau$ are learned by backpropagation. Optional learned pieces are a small U-Net used as a regularizer inside the iterations (Le-ADMM*) and a larger U-Net applied as a final denoiser (Le-ADMM-U). The learned parameters are what absorb forward-model error, while the fixed update structure is what keeps the network bounded-compute and sample-efficient.","core_discovery":"The central claim is that unrolling ADMM with the lensless-specific variable splitting $b=CHx$ into five layers produces reconstructions comparable to 100-iteration ADMM at one twentieth of the runtime, and that adding learned components lets the network absorb model mismatch rather than being penalized by it. Concretely, Le-ADMM-U averages LPIPS $0.1904$ and MSE $0.0074$ on a 1,000-image held-out set, versus $0.5711$ and $0.0622$ for converged ADMM, in 75 ms rather than 1.5 s. The paper also claims that this training transfers beyond the flat-screen training setup: its networks reconstruct natural objects in the wild with image quality similar to or better than converged ADMM.","pith_inferences":["The reported 20× speedup is for GPU execution at a fixed five-layer budget; extending interactive preview to low-power or mobile hardware would require quantization or distillation, which the paper does not address.","A natural testable extension is to train the same unrolled architecture on multiple depth-dependent PSFs, which could turn the current flat-screen-trained network into a depth-aware lensless imager without changing the unrolling principle.","Because the lensed-camera labels define what 'ground truth' means, the achievable quality is bounded by the training camera's aberrations and field of view; applications without a lensed reference would need a self-supervised or physics-consistency loss.","The observed data-fidelity-versus-quality trade-off suggests that a practical product could expose a 'fidelity slider' by varying the number of unrolled layers or the denoiser strength."],"forward_implications":["A DiffuserCam-style lensless imager using Le-ADMM or Le-ADMM-U can show the user a reconstruction in about 75 ms, making interactive preview of the scene feasible rather than waiting roughly 1.5 s for converged ADMM.","On the 1,000-image test set, Le-ADMM-U improves average LPIPS from 0.5711 to 0.1904 and MSE from 0.0622 to 0.0074, so perceptual quality is about three times closer to the lensed-camera ground truth.","With only 20 learned parameters, Le-ADMM matches the perceptual quality of converged ADMM while using far fewer training pairs, which suggests the physical model does most of the work.","Because the learned networks trade data fidelity for perceptual quality, users can choose where to sit on the model-to-deep spectrum depending on whether measurement consistency or visual appeal matters more.","The same networks, trained on images of a flat screen, reconstruct natural 3D objects under different lighting in the paper's qualitative wild tests, indicating the approach is not confined to the training display."],"supporting_citations":[{"why":"Defines the DiffuserCam forward model $b=CHx$ and the ADMM variable splitting that the unrolled networks are built from.","marker":"[3]"},{"why":"Introduces the DiffuserCam prototype and PSF-calibration procedure used for experimental measurements and as the classic reconstruction baseline.","marker":"[12]"},{"why":"Provides the ADMM algorithm and update equations that are unrolled into trainable network layers.","marker":"[15]"},{"why":"Establishes the technique of interpreting a fixed number of optimization iterations as a trainable network, which the paper adapts to lensless imaging.","marker":"[19]"},{"why":"Supplies the U-Net architecture used as the learned regularizer in Le-ADMM*, as the final denoiser in Le-ADMM-U, and as the purely deep comparison.","marker":"[24]"},{"why":"Defines the LPIPS perceptual metric used both as the training loss component and as the evaluation measure for image quality.","marker":"[25]"},{"why":"Provides the collection of natural images displayed on the monitor to form the 25,000 aligned lensed/lensless training pairs.","marker":"[27]"}],"fun_headline_variants":["Lensless imaging in 75 ms: unrolled ADMM beats converged","Unrolled ADMM reconstructs lensless images 20× faster","Physics-aware network speeds lensless cameras 20×","Neural unrolling improves lensless image quality and speed","Learned unrolling makes lensless imaging practical"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a single measured point-spread function, taken at the training distance, describes how the camera blurs any scene point, including natural scenes at other depths and with different lighting; if that shift-invariance breaks, the wild-scene reconstructions do not establish generalization.","fun_headline_variants_meta":{"raw":{"variants":["Lensless imaging in 75 ms: unrolled ADMM beats converged","Unrolled ADMM reconstructs lensless images 20× faster","Physics-aware network speeds lensless cameras 20×","Neural unrolling improves lensless image quality and speed","Learned unrolling makes lensless imaging practical"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000983,"raw_usage":{"total_tokens":4155,"prompt_tokens":914,"completion_tokens":3241,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":530,"completion_tokens_details":{"reasoning_tokens":3156}},"tokens_in":530,"tokens_out":3241,"duration_ms":22837,"temperature":1.0,"reasoning_tokens":3156,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T10:12:52.733685+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure a point source with the same DiffuserCam at several distances, such as 10 cm, 50 cm, and 2 m, and compare the recorded PSFs: if the pattern changes shape with depth instead of simply translating, the shift-invariant forward model in Eq. (1) is violated. A complementary test is to collect aligned lensed/lensless pairs of natural 3D scenes and compute LPIPS and MSE for Le-ADMM-U versus converged ADMM; if the learned advantage disappears outside the monitor plane, the generalization claim is not supported.","supporting_citations":[{"cited_title":"DiffuserCam: lenslesssingle-exposure 3D imaging,","cited_arxiv_id":null,"evidence_quote":"Defines the DiffuserCam forward model $b=CHx$ and the ADMM variable splitting that the unrolled networks are built from."},{"cited_title":"DiffuserCam: diffuser-based lensless cameras,","cited_arxiv_id":null,"evidence_quote":"Introduces the DiffuserCam prototype and PSF-calibration procedure used for experimental measurements and as the classic reconstruction baseline."},{"cited_title":"Distributed optimization and statistical learning via the alternating direction method of multipliers,","cited_arxiv_id":null,"evidence_quote":"Provides the ADMM algorithm and update equations that are unrolled into trainable network layers."},{"cited_title":"Learning fast approximations of sparse coding,","cited_arxiv_id":null,"evidence_quote":"Establishes the technique of interpreting a fixed number of optimization iterations as a trainable network, which the paper adapts to lensless imaging."},{"cited_title":"U-net: Convolutional networks for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net architecture used as the learned regularizer in Le-ADMM*, as the final denoiser in Le-ADMM-U, and as the purely deep comparison."},{"cited_title":"The unreasonable effectiveness of deep features as a perceptual metric,","cited_arxiv_id":null,"evidence_quote":"Defines the LPIPS perceptual metric used both as the training loss component and as the evaluation measure for image quality."},{"cited_title":"The MIR Flickr Retrieval Evaluation,","cited_arxiv_id":null,"evidence_quote":"Provides the collection of natural images displayed on the monitor to form the 25,000 aligned lensed/lensless training pairs."}],"review_version":1}