{"id":"6c1d44bb-415d-4fe3-97a9-55fc2b186445","arxiv_id":"2501.03511","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new two-stage pipeline for low-light lensless imaging that uses Wiener filtering for a first estimate and a wavelet-domain conditional diffusion model for denoising and brightness enhancement.","lead":"A two-stage reconstruction method, combining a learnable Wiener filter with a wavelet-domain conditional diffusion model, produces brighter and less noisy lensless images in low light. The approach beats several existing reconstruction methods on simulated and real camera data, though not on every metric.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Real-world validation may overlap with training: the paper's 900/100 simulated split and its statement that measured targets included 'all target images in the original dataset' leave open the possibility that the measured test set contains images already seen during training, inflating Tables 4…","rationale":"The reader's weakest assumption (PSF/noise fidelity) is a legitimate generalization risk, but the paper partially addresses it by building the actual camera and using real measurements. The more acute risk is that the real-world validation may not be a held-out test at all: the described capture of 'all target images in the original dataset' creates an unresolved overlap with the 900 simulated training images. If that overlap exists, the diffusion module's ability to reproduce previously seen clean images would inflate the primary quantitative evidence, independent of any forward-model mismatch. I therefore highlight this split ambiguity as the single most load-bearing concern. The current conditional verdict remains appropriate because the issue is checkable and could be resolved by releasing the split and rerunning the measured evaluation; I would not reject the paper outright without confirming the overlap. Secondary issues noted in the reader's verdict, such as the 'outperforms all other methods across all metrics' claim being contradicted by Table 1 and Table 3 SSIM values, further support the conditional stance but are not the central load-bearing concern.","tokens_in":13308,"tokens_out":9938,"duration_ms":95582,"concrete_test":"Obtain the exact list of LOLv2 image indices projected for the measured dataset and compare it with the 900 training indices. Re-run the measured-data evaluation using only the 100 held-out test images, reporting per-image metrics; if the current Table 4/5 results include any training images, recompute them on the disjoint held-out subset and assess whether the reported advantage over MWDN and DeepLIR persists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central evidence for 'substantial improvement' is the measured-data evaluation in Section 4.5. Section 4.1 divides 1000 LOLv2 pairs into 900 training and 100 test pairs for simulated data, then says the real-world validation used the LOLv2 Synthetic subset for projection. Section 4.5 describes switching target images 'until all target images in the original dataset have been traversed.' The manuscript never states that the real-world targets are exactly the 100 held-out test images, nor that they are disjoint from the 900 training images. If the 900 training clean images were among the projected targets, the conditional diffusion model has already seen the exact ground-truth images during training, making the Table 4 real-world metrics and Table 5 exposure sweep optimistic and the comparison against non-trained baselines unfair. Since no code or data are released, this cannot be checked from the manuscript. This is more directly load-bearing than PSF/noise-model mismatch: even a perfect forward model cannot validate a method on data that are not held out.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage reconstruction method for lensless imaging in low-light conditions. The first stage applies a learnable Wiener filter based on a calibrated point-spread function to obtain an initial noisy reconstruction; the second stage refines it with a conditional diffusion model operating in a wavelet domain, separating low-frequency and high-frequency processing. The method is evaluated on simulated data derived from the LOLv2 dataset and on real measurements from a self-built lensless camera, with comparisons against ADMM, U-Net, FlatNet, MWDN, and DeepLIR. The authors claim substantial improvements in brightness, noise suppression, and sharpness, and state that the proposed method outperforms all baselines across all metrics.","tokens_in":13532,"tokens_out":4095,"duration_ms":38267,"significance":"If the claims are sustained, the work is a useful contribution to low-light lensless imaging: it combines a physics-based first stage with a generative refinement stage, includes a systematic camera noise model, and validates on a real hardware prototype. The explicit treatment of low-light noise, the wavelet-domain diffusion design, and the comparison against several recent baselines are strengths. However, the evaluation is currently undermined by an unsupported 'all metrics' claim, missing statistical characterization, and an unclarified relationship between the training set and the real-world test targets. The absence of code or data makes these points unverifiable from the manuscript alone.","major_comments":[{"comment":"The text states that the proposed method 'outperforms the others across all metrics,' but the authors' own tables contradict this. In Table 1, MWDN achieves higher SSIM (0.6115) than the proposed method (0.5719); in Table 3, MWDN again has higher SSIM (0.5275 vs. 0.4951). The claim must be revised to identify the specific metrics on which the method improves, and the SSIM deficit should be discussed rather than ignored.","section":"§4.4, Tables 1 and 3"},{"comment":"The real-world validation does not establish that the measured target images are disjoint from the training set. Section 4.1 states that 1000 LOLv2 pairs are split into 900 training and 100 test pairs for simulation, and that the real-world validation uses the LOLv2 Synthetic subset. Section 4.5 then says target images are switched 'until all target images in the original dataset have been traversed' without specifying that these targets are exactly the 100 held-out test images or that they are not among the 900 training images. If training clean images were among the projected targets, the conditional diffusion model has already seen the ground-truth content, which would inflate the real-world results in Table 4 and Table 5 and make the comparison against non-trained baselines unfair. The manuscript must specify the exact target set and confirm its disjointness from the training set.","section":"§4.1 and §4.5"},{"comment":"All quantitative results are reported as averages without error bars, confidence intervals, or the number of images per entry. Several between-method differences are small (e.g., PSNR differences around 0.2 dB), and without statistical characterization it is not possible to determine whether these differences are meaningful. Please provide per-image score distributions, error bars, or a statistical significance test, and state how many images each average covers.","section":"Tables 1, 3, 4, and 5"},{"comment":"The method's first stage relies on the calibrated PSF in the forward model b = Hx, and the paper itself notes that model mismatch is a known issue. The experiments do not quantify sensitivity to PSF calibration error or to deviations of the real camera noise from the simulation parameters in Table 2. Because the diffusion model is trained on simulated data generated with this forward model, a robustness test with perturbed PSFs or with an independent noise parameter setting would strengthen the claim that the method works under realistic model mismatch.","section":"§3.1, Eq. (6), and §4.5"}],"minor_comments":[{"comment":"The heading contains a typo: 'Fisrt Stage' should be 'First Stage'.","section":"§3.1"},{"comment":"The displayed formula has malformed parentheses: 'p_θ(xt−1)| xt, s)' should likely be 'p_θ(xt−1 | xt, s)'.","section":"§3.2.1, Eq. (12)"},{"comment":"The sensor specifications are inconsistent across sections: §4.2 reports an IMX219 sensor with 1.12 µm pixel size and 2028×1520×4 Bayer input, §4.4 reports a pixel pitch of 0.014 mm, and §4.5 reports a raw resolution of 4056×3040. Please clarify the actual sensor geometry and the relationship between these numbers.","section":"§4.2, §4.4, and §4.5"},{"comment":"All three table captions say 'simulation test set,' but Table 4 reports measured real-world data and Table 5 reports varying-exposure measured data. The captions should be corrected.","section":"Tables 3, 4, and 5"},{"comment":"The U-Net row consists of dashes with no explanation. If U-Net failed to train or was excluded, this should be stated explicitly; otherwise the reader cannot interpret the comparison.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal and the core idea is plausible. The main obstacle is not the method itself but the evaluation: the 'all metrics' statement is internally contradicted, and the real-world test set may overlap with training data. If the authors can clarify the data split, provide statistical support, and revise the overclaims, the paper could become acceptable. I would not reject based solely on the current ambiguity, but the ambiguity is load-bearing and must be resolved."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"What is actually new: a two-stage low-light lensless reconstruction pipeline that pairs a learnable Wiener filter with a wavelet-domain conditional diffusion model, plus bidirectional training for stability. The noise-model analysis in Section 2 is a solid, practical contribution, and building a real lensless camera and testing on measured data is genuinely valuable. The exposure sweep (0.3, 0.5, 0.7 s) is a nice robustness check. On simulation and measured data, the method shows consistent gains in PSNR, LPIPS, and MSE over ADMM, FlatNet, and DeepLIR, and the visual examples look better.\n\nThe soft spots, in proportion. The biggest one is the stress-test concern, and it lands. The paper splits 1000 LOLv2 pairs into 900 training and 100 test for simulation, but for the real-world validation it says the LCD displayed \"all target images in the original dataset\" (Section 4.5). It never states that the measured targets are the held-out 100, nor that they are disjoint from the 900 training clean images. If any of the training clean images were projected and measured, the conditional diffusion model has already seen the exact ground-truth images, making the Table 4 metrics and the Table 5 exposure sweep optimistic and the comparison against baselines unfair. No code or data are released, so this cannot be checked. This is load-bearing, not a minor caveat.\n\nSecond, the claim in Section 4.4 that the method \"outperforms all other methods across all metrics\" is contradicted by the paper's own Tables 1 and 3: MWDN has higher SSIM in both (0.6115 vs 0.5719; 0.5275 vs 0.4951). That overclaim should be corrected. Third, there are no error bars, and the U-Net baseline appears as dashes without explanation. Minor-to-moderate reporting issues. Fourth, the real-world experiments are limited to LCD-displayed scenes in a controlled lab, which the Discussion honestly acknowledges; that is a scope limitation, not a fatal flaw.\n\nWho this is for: anyone working on lensless imaging or low-light computational imaging, especially two-stage physics-plus-generation methods. The engineering is thoughtful and the idea is worth engaging with. But the real-world validation needs to be re-done or re-clarified on a truly held-out measured set, and the metric claims need to match the tables.\n\nRecommendation: this deserves a serious referee, not a desk rejection. My own verdict would be major revision: clarify/guarantee the measured test set is disjoint from training, fix the overclaim, add error bars, and ideally release code and data so the measured results can be audited.","headline":"A genuinely useful two-stage lensless low-light reconstruction method, with real hardware experiments, but the measured-data validation may have reused training images and the 'outperforms all metrics' claim is contradicted by the paper's own tables.","tokens_in":14042,"tokens_out":2225,"would_cite":false,"duration_ms":22168,"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":"A two-stage pipeline combining a learnable Wiener filter with a wavelet-domain diffusion model can reconstruct clear, bright images from low-light lensless measurements.","keywords":["lensless imaging","low-light imaging","conditional diffusion model","Wiener filter","wavelet transform","computational imaging","image reconstruction","generative model"],"falsifier":"Measure the reconstruction quality of a model trained with the paper's simulated noise when the real camera is operated at a different temperature, with a different mask-sensor distance, or with a scene containing depth variation; if the PSNR/SSIM margin over simple Wiener deconvolution shrinks to roughly zero, the calibrated-PSF and simulated-noise assumptions are the load-bearing parts.","tokens_in":13129,"feed_emoji":"📷","tokens_out":4632,"duration_ms":39263,"temperature":0.7,"pith_summary":"This paper is trying to establish that lensless cameras—which capture light through a mask instead of a lens—can still produce high-quality images in dim environments, a regime where existing reconstruction methods fail. The proposed answer is a two-stage pipeline: a physics-based learnable Wiener filter first turns raw noisy measurements into a rough, noisy image, then a conditional diffusion model operating in the wavelet domain removes the noise, restores brightness, and fills in missing detail. The authors argue that this combination of model-driven first stage and generative second stage is what makes the method robust under photon-starved conditions, and they support it with simulations and a self-built camera experiment. If true, it would make lightweight, lensless imagers practical in low-light settings such as space or night-time inspection.","feed_headline":"Two-stage model sharpens low-light lensless images","feed_subtitle":"A learnable Wiener filter plus wavelet-domain diffusion restores brightness and detail from dim, noisy captures.","key_machinery":"The load-bearing mechanism is the two-stage reconstruction with four linked components. First, the learnable Wiener filter implements the pseudo-inverse of the forward model $\\mathbf{b} = \\mathbf{H}\\mathbf{x}$ in the frequency domain, with the PSF $\\mathbf{h}$ initialized from calibration and treated as learnable, producing an initial image $\\hat{x}$. Second, the Haar discrete wavelet transform splits this image into low-frequency (LL) and high-frequency (LH, HL, HH) sub-bands, reducing spatial resolution fourfold and separating brightness/noise from texture. Third, a conditional diffusion model, trained with both forward diffusion and reverse denoising (bidirectional training), is conditioned on the LL sub-band to generate a clean, bright low-frequency component. Fourth, a depthwise-separable convolutional network with cross-attention between the high-frequency sub-bands restores fine details. The loss combines noise-prediction MSE, reconstruction MAE+SSIM+LPIPS, and a high-frequency MSE+TV term.","core_discovery":"The central claim is that low-light lensless reconstruction is best tackled by separating the problem into two guided steps rather than trying to invert the noisy forward model in one go. The first step uses the known point-spread function to project measurements into the range space of the adjoint operator via a learnable Wiener filter, yielding an initial reconstruction that is noisy and dim but encodes the image content. The second step refines this in the wavelet domain: a conditional diffusion model cleans the low-frequency sub-band (where brightness and noise live) while a depthwise-separable convolutional network with cross-attention restores high-frequency texture. The authors report that this method outperforms both traditional optimization (ADMM) and learned baselines (U-Net, FlatNet, MWDN, DeepLIR) under simulated and real low-light conditions, with particular gains in brightness, noise suppression, and perceptual quality.","pith_inferences":["A natural extension is to replace the fixed calibrated PSF with a PSF that is jointly estimated or made uncertainty-aware; if the first stage becomes robust to PSF error, the same pipeline could apply to out-of-focus or thermal-drift conditions without retraining.","The wavelet-domain conditioning could be generalized to other inverse problems with strong noise, such as low-light phase retrieval or holography, since the brightness/noise separation is not specific to lensless optics.","The paper's simulated dataset only covers images displayed on an LCD screen; a harder test would be natural scenes with depth variation, where the shift-invariant PSF assumption is violated, and the method may need a spatially-variant extension."],"forward_implications":["If the method is correct, a lensless camera with a simple amplitude mask and a modest CMOS sensor can operate at exposure times of 0.3 to 0.7 seconds and still yield images close to ground truth in brightness and detail.","The two-stage design implies that the same diffusion refinement can be reused for different mask designs or sensor noise levels, as long as the first-stage Wiener filter is re-calibrated to the new PSF.","The wavelet-domain decomposition reduces memory and computation enough that the diffusion process runs at roughly 0.4 seconds per image, making the approach plausible for field or near-real-time deployments.","The method's robustness across exposure times suggests that the generative second stage, not the raw measurement SNR, is the main driver of final image quality."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model formulation (forward diffusion, reverse generation, noise prediction) that the second stage adapts into a conditional, wavelet-domain version.","marker":"[27]"},{"why":"Supplies the wavelet-domain diffusion design for low-light image enhancement, from which the paper borrows the Haar DWT decomposition and conditioning strategy.","marker":"[28]"},{"why":"Supplies the LOLv2 dataset of low/normal-light image pairs used to synthesize the low-light lensless training and test data.","marker":"[31]"},{"why":"Supplies the learnable Tikhonov/Wiener-style first-stage reconstruction idea and is the FlatNet baseline that the method must beat.","marker":"[18]"},{"why":"Supplies the DiffuserCam dataset for phase-mask generalization experiments and the U-Net baseline architecture.","marker":"[17]"},{"why":"Supplies the model-mismatch analysis and multi-stage compensation idea that motivates the two-stage design.","marker":"[19]"}],"fun_headline_variants":["Wavelet diffusion clears dim lensless shots","Wiener-plus-diffusion revives low-light lenses","Two-step AI lifts lensless images from the dark","Low-light lensless crispness via dual guidance","Generative two-stage fix for noisy lensless views"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that the calibrated point-spread function and the simulated camera-noise model match the real sensor's behaviour; if those disagree with the actual optics and electronics, both the Wiener first stage and the diffusion model trained on simulated data will degrade.","fun_headline_variants_meta":{"raw":{"variants":["Wavelet diffusion clears dim lensless shots","Wiener-plus-diffusion revives low-light lenses","Two-step AI lifts lensless images from the dark","Low-light lensless crispness via dual guidance","Generative two-stage fix for noisy lensless views"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000118,"raw_usage":{"total_tokens":1073,"prompt_tokens":924,"completion_tokens":149,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":540,"completion_tokens_details":{"reasoning_tokens":75}},"tokens_in":540,"tokens_out":149,"duration_ms":2402,"temperature":1.0,"reasoning_tokens":75,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:51:37.430113+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the reconstruction quality of a model trained with the paper's simulated noise when the real camera is operated at a different temperature, with a different mask-sensor distance, or with a scene containing depth variation; if the PSNR/SSIM margin over simple Wiener deconvolution shrinks to roughly zero, the calibrated-PSF and simulated-noise assumptions are the load-bearing parts.","supporting_citations":[{"cited_title":"Denoisingdiffusionprobabilisticmodels,","cited_arxiv_id":null,"evidence_quote":"Supplies the denoising diffusion probabilistic model formulation (forward diffusion, reverse generation, noise prediction) that the second stage adapts into a conditional, wavelet-domain version."},{"cited_title":"Low-light image enhancement with wavelet-based diffusion models,","cited_arxiv_id":null,"evidence_quote":"Supplies the wavelet-domain diffusion design for low-light image enhancement, from which the paper borrows the Haar DWT decomposition and conditioning strategy."},{"cited_title":"Sparse gradient regularized deep retinex network for robust low-light image enhancement,","cited_arxiv_id":null,"evidence_quote":"Supplies the LOLv2 dataset of low/normal-light image pairs used to synthesize the low-light lensless training and test data."},{"cited_title":"Flatnet: Towards photorealistic scene reconstruction from lensless measurements,","cited_arxiv_id":null,"evidence_quote":"Supplies the learnable Tikhonov/Wiener-style first-stage reconstruction idea and is the FlatNet baseline that the method must beat."},{"cited_title":"Learned reconstructions for practical mask-based lensless imaging,","cited_arxiv_id":null,"evidence_quote":"Supplies the DiffuserCam dataset for phase-mask generalization experiments and the U-Net baseline architecture."},{"cited_title":"Robust reconstruction with deep learning to handle model mismatch in lensless imaging,","cited_arxiv_id":null,"evidence_quote":"Supplies the model-mismatch analysis and multi-stage compensation idea that motivates the two-stage design."}],"review_version":1}