{"id":"3f93fc38-6bb2-424a-99ca-20f844b57f91","arxiv_id":"2607.04608","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Interleaving differentiable forward projections with learnable inverse updates at every encoder–decoder scale yields state-of-the-art lensless image reconstruction and transfers to deblurring and holography.","lead":"IFIN is a neural reconstructor for lensless cameras that repeatedly runs the camera’s physics forward and backward inside every layer of a multi-scale network. It sets a new accuracy bar on three hard lensless benchmarks and works on related blur and holography tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper’s strongest claim is empirical and architectural, not a universal guarantee under arbitrary model mismatch. Ablations (Table 4 / S1) show that removing either direction of the FSO/ISO coupling or replacing both with identities substantially hurts metrics, so the bidirectional design is load-bearing for the gains rather than incidental capacity. Learned PSF fields improve over frozen calibrated kernels on MWNet, and the holography experiment shows the interleaving template is not tied to a single operator family. Limitations (FFT cost scaling with k, paired supervision, failure under saturation/geometry change) are disclosed and do not contradict the results under the evaluated conditions. The reader’s weakest assumption correctly flags the standard physics-model prerequisite; it is not a hidden flaw that overturns the evidence. Therefore the ACCEPT / high-confidence verdict should remain unchanged.","tokens_in":25920,"tokens_out":493,"duration_ms":5581,"concrete_test":"Independently retrain IFIN (k=1 and k=9) and the two strongest baselines (UPDN, MoDL) on DiffuserCam and MWNet with identical loss, data splits, and evaluation crops; if the reported PSNR margins (+1.63 dB / +2.58 dB) shrink below ~0.5 dB or reverse, the SOTA claim would need qualification. Otherwise the claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that bidirectional FSO/ISO interleaving at every encoder–decoder scale, with a jointly learned shift-variant PSF field, yields SOTA lensless reconstruction and transfers to related inverse problems. The manuscript supports this with three benchmarks (including a new WiderCam set), strong baselines, component ablations isolating the bidirectional coupling (identity FSO/ISO, FSO-only, ISO-only), learned-vs-calibrated PSF comparisons, and a holography transfer experiment that swaps operators while keeping the interleaving backbone. The reader’s weakest assumption (need for a sufficiently faithful differentiable linear forward model and target-domain paired data) is real but is standard for hybrid physics–NN reconstruction and is explicitly acknowledged by the authors (Sec. 6, S12); it does not undermine the reported gains under the stated operating regime. No internal inconsistency or unsupported leap in the strongest claim is evident from the provided evidence.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes IFIN, an encoder–decoder architecture for lensless image reconstruction that interleaves differentiable forward system operators (FSO) with learnable inverse system operators (ISO) inside Integrated Forward–Inverse Blocks at every scale, while jointly learning a multi-kernel shift-variant PSF field with ROI blending. Motivated by locality mismatch, one-sided inversion, and representation bottlenecks under large-footprint PSFs, the design keeps coupled measurement- and image-domain streams so that physics consistency cues are reintroduced throughout decoding. Empirically, IFIN reports state-of-the-art PSNR/LPIPS/SSIM on DiffuserCam, a newly introduced wide-FoV WiderCam dataset, and MultiWienerNet, with component ablations isolating bidirectional coupling and learned vs. calibrated PSFs, plus transfer experiments on Gaussian deblurring and simulated inline holography (with modality-matched operators).","tokens_in":26220,"tokens_out":1303,"duration_ms":17734,"significance":"If the reported gains hold under fair comparison, the work is a solid contribution to hybrid physics–NN reconstruction for computational imaging. The bidirectional multi-scale interleaving is a clear architectural idea relative to one-sided unrolled or single-inversion hybrids, the learnable PSF field reduces dense calibration burden, and the new WiderCam benchmark (strong shift variance, >100° FoV) is a useful community resource. Code, supplementary material, and the WiderCam dataset are promised, which strengthens reproducibility. The holography transfer, while simulated, supports the claim that the interleaving template is not tied to a single convolutional PSF model.","major_comments":[{"comment":"Table 1 SOTA margins, especially WiderCam (+0.65 dB PSNR over MoDL) and several LPIPS/SSIM gaps, are reported from single runs with no error bars, multi-seed statistics, or significance tests. Given free parameters (k, loss weights, gates, ϵ, ROI maps) and the claim of state-of-the-art quality, please report mean±std over at least 3 seeds for IFIN and the strongest baselines on all three lensless benchmarks, or otherwise justify that the ranking is stable.","section":null},{"comment":"§5 and S1–S2: fairness of the hybrid baseline comparison is not fully transparent. The text states a “unified reconstruction objective” and fixed (λ_img, λ_perc) for retrained learning-based methods, but does not tabulate whether each baseline (UPDN, MWDNs, LensNet, MoDL, MWNet, etc.) was retrained end-to-end on the same splits/resolutions with matched epochs, PSF normalization, and data augmentation, or evaluated from published checkpoints. Please add a short protocol table so the +1.63/+0.65/+2.58 dB claims can be audited.","section":null},{"comment":"§4.2 and S9: the multi-kernel design is asymmetric—FSO uses a single averaged PSF while ISO applies region-wise Wiener-like inversion with ROI blending. The efficiency argument is reasonable, but the paper does not quantify how much accuracy is left on the table versus a shift-variant FSO (even on a subset). A controlled experiment or bound on the measurement-consistency residual under the averaged FSO would strengthen the claim that bidirectional coupling, not only multi-kernel ISO, drives the gains under strong shift variance (WiderCam/MWNet).","section":null}],"minor_comments":[{"comment":"Table 5 / S11: inference latency for IFIN (k=1) is ~142 ms vs. 12–60 ms for strong baselines; the conclusion correctly flags FFT overhead, but a brief note on whether mixed-precision, reduced padding, or fewer IFIB stages can close the gap would help practitioners choose the k=1 operating point.","section":null},{"comment":"Fig. 1 notation: n indexes both the per-scale PSF embedding h_n and stage-(n) representations; a one-line clarification that h_n is shared across encoder/decoder IFIBs at the same scale would reduce ambiguity.","section":null},{"comment":"Eq. (4): ϵ(u,v) is described as a learnable non-negative 2D regularizer; state its spatial resolution relative to the FFT grid and whether it is shared across scales or per IFIB.","section":null},{"comment":"WiderCam (S4): affine alignment is estimated offline from a deconvolution baseline and LoFTR correspondences. Please state whether any test images were excluded for failed registration and whether metrics are computed after inverse-warping only (as claimed) without additional cropping that could favor methods with less geometric distortion.","section":null},{"comment":"Table 2: for σ=5, IFIN trails NAFNet/RCAN slightly; the text already notes graceful degradation for larger blur, but a short clause that the physics prior is not free in the mild-blur regime would set expectations.","section":null},{"comment":"Inline holography (Tab. 3, S5) is fully simulated; the abstract’s “suggesting … can extend beyond lensless cameras” is fair, but please mark the holography result as simulated in the abstract or introduction to avoid over-reading experimental transfer.","section":null},{"comment":"Typos/style: abstract and title use both “Forward-Inverse” and “Forward–Inverse”; unify en-dash usage. Occasional missing spaces after periods in the compiled text (e.g., “reconstructionfromtheresulting”) appear to be PDF extraction artifacts—verify the camera-ready source.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The work is a strong empirical systems paper for a computational imaging / CV venue. I do not see a load-bearing correctness failure; the main risks are incomplete multi-seed reporting and incomplete documentation of baseline retraining. If the authors supply those, accept is appropriate. WiderCam is a genuine plus if released as promised."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a clean, usable advance in hybrid physics–ML reconstruction for lensless cameras. The new piece is not “add a Wiener layer,” but the systematic dual-stream interleaving of a differentiable forward operator and a learnable inverse at every encoder–decoder scale, plus a jointly learned multi-kernel PSF field with ROI blending. That design is carried through carefully and produces clear gains.\n\nWhat they do well: three real benchmarks (DiffuserCam, their new WiderCam with strong field dependence, and MultiWienerNet), strong classical/CNN/hybrid baselines, and ablations that actually isolate the claim—identity operators, FSO-only, ISO-only, regularizer, padding, k-scaling, learned vs frozen PSFs. The holography transfer (swap angular-spectrum operators, keep the interleaving backbone) is a useful check that the idea is not locked to convolution. Code, supplement, and WiderCam are promised; the modeling assumptions (linear differentiable forward model, paired target-domain data) are standard for the field and are stated honestly in the limitations.\n\nSoft spots are real but secondary. FFT-based FSO/ISO cost scales with k and padding; k=1 already beats the baselines, so denser fields are an accuracy–compute trade-off rather than a free lunch. No error bars or multi-seed stats. Gains on Gaussian deblurring are modest until the blur is large. None of that undercuts the main numbers (+1.63 / +0.65 / +2.58 dB PSNR on the three lensless sets).\n\nThis is for people building or deploying thin computational cameras and for anyone designing multi-scale physics–NN invertors. The math is standard Fourier deconvolution with learnable regularizers; the data and citation pattern look solid. I would send it to referees and would cite the architecture and WiderCam if I work in this space.","headline":"Solid hybrid architecture for lensless reconstruction: multi-scale bidirectional FSO/ISO plus a learned PSF field, with real SOTA gains and a new dataset.","tokens_in":26793,"tokens_out":468,"would_cite":true,"duration_ms":4882,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Interleaving forward projections with inverse updates at every scale recovers sharp images from highly multiplexed lensless measurements and adapts the point-spread field under mismatch.","keywords":["lensless imaging","computational imaging","inverse problems","model-based deep learning","point-spread function","shift-variant deconvolution","encoder-decoder"],"falsifier":"Train and evaluate the identical architecture with the forward and inverse operators replaced by identity maps (or with a deliberately mismatched fixed PSF) on DiffuserCam or WiderCam; if the reported PSNR/LPIPS gains over strong hybrid baselines disappear, the claim that bidirectional physics coupling is the source of the improvement is falsified.","tokens_in":26850,"feed_emoji":"📷","tokens_out":669,"duration_ms":5440,"temperature":0.7,"pith_summary":"Lensless cameras replace bulky lenses with thin coded elements, but the resulting measurements mix light over large spatial footprints, so inversion is severely ill-conditioned and fragile to calibration error. Existing hybrid networks often apply physics only once, after which measurement cues are lost. This paper claims that repeatedly coupling a differentiable forward operator with a learnable inverse operator inside every level of an encoder–decoder hierarchy keeps measurement-domain and image-domain information available to each other throughout refinement. The same shared learnable point-spread field drives both operators, so the kernels themselves adapt under model uncertainty. On three lensless benchmarks, including a new wide-field dataset, the method reports clear gains in reconstruction quality; the same interleaving pattern also remains competitive when the operators are swapped for Gaussian deblurring or for angular-spectrum holography. The result matters because it turns physics consistency into a continuous architectural principle rather than a single pre- or post-processing step, and because it reduces the need for dense per-system calibration.","feed_headline":"Forward-inverse blocks recover sharp lensless images","feed_subtitle":"Physics operators at every scale lift PSNR by up to 2.6 dB and adapt the PSF under mismatch","key_machinery":"The Integrated Forward–Inverse Block (IFIB): at each scale it runs a Forward System Operator that projects the current image features into the measurement domain and an Inverse System Operator that restores image features from the measurement, then fuses both streams with learnable gates before the next scale, all driven by a shared multi-scale point-spread field.","core_discovery":"The authors establish that an encoder–decoder whose every scale interleaves a differentiable forward projection with a learnable inverse update, both conditioned on a jointly learned shift-variant point-spread field, yields state-of-the-art reconstruction quality on DiffuserCam, WiderCam and MultiWienerNet and remains competitive when the same interleaving is applied to Gaussian deblurring and simulated inline holography.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["IFIN interleaves forward-inverse steps for sharper lensless recovery","Physics operators at all scales recover lensless images under mismatch","Forward-inverse interleaving adapts PSFs for state-of-the-art reconstruction","Bidirectional network refines lensless images via joint measurement-image cues","Scale-wise forward projections lift lensless quality with learned PSF fields"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"A sufficiently faithful, differentiable linear forward model (and a compatible inverse or adjoint) must still be available at test time; severe saturation or large geometry changes erase recoverable information that no reconstructor can restore.","fun_headline_variants_meta":{"raw":{"variants":["IFIN interleaves forward-inverse steps for sharper lensless recovery","Physics operators at all scales recover lensless images under mismatch","Forward-inverse interleaving adapts PSFs for state-of-the-art reconstruction","Bidirectional network refines lensless images via joint measurement-image cues","Scale-wise forward projections lift lensless quality with learned PSF fields"]},"model":"grok-4.5","effort":"low","cost_usd":0.004194,"raw_usage":{"total_tokens":1225,"prompt_tokens":741,"num_sources_used":0,"completion_tokens":97,"cost_in_usd_ticks":41940000,"prompt_tokens_details":{"text_tokens":741,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":387,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":741,"tokens_out":97,"duration_ms":6544,"temperature":1.0,"reasoning_tokens":387,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T16:29:35.294389+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train and evaluate the identical architecture with the forward and inverse operators replaced by identity maps (or with a deliberately mismatched fixed PSF) on DiffuserCam or WiderCam; if the reported PSNR/LPIPS gains over strong hybrid baselines disappear, the claim that bidirectional physics coupling is the source of the improvement is falsified.","supporting_citations":[],"review_version":1}