{"id":"2710edd0-9008-4422-a216-891eb9e822a7","arxiv_id":"2411.15255","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"OSMamba, a Mamba-based exposure correction network with omnidirectional spectral scanning and a dual-domain diffusion prior generator, reports state-of-the-art PSNR/SSIM on MSEC, SICE, and LCDP.","lead":"Researchers built a neural network that corrects underexposed and overexposed photos by scanning frequency-domain information in multiple directions and using a learned prior generator to restore lost details. It reports state-of-the-art scores on three benchmark datasets, which matters for camera pipelines, photo editing, and downstream vision tasks.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim depends on the DDPG prior staying informative when initialized from pure noise at inference; the only evidence is a single LCDP ablation with no error bars, and a train/inference diffusion-start mismatch is not analyzed.","rationale":"The paper is a coherent engineering contribution. The OS-SSM design is reasonable, and the ablations in Tables 2 and 3 support the spectral-scanning contribution, at least on LCDP. The reported gains are plausible, and I found no internal inconsistency in the equations. My concern is not that the authors are wrong, but that the most novel load-bearing component, the GT-free DDPG, is supported by a single ablation that does not isolate the diffusion mechanism and has a specific train/inference distribution mismatch. The central SOTA claim on LCDP is quantitatively dependent on this module, so the evidence bar should be higher. A dual-domain extractor-only control and a pure-noise versus teacher-conditional start comparison would settle the issue. This keeps the reader's CONDITIONAL verdict unchanged.","tokens_in":16067,"tokens_out":9016,"duration_ms":96997,"concrete_test":"Run the decisive control on LCDP: (i) add a dual-domain extractor-only variant (DDPE* without the denoiser, prior produced directly from D) and (ii) for the full model, generate priors from both N(0,I) (inference) and N(sqrt(bar-alpha_T) Z, (1 - bar-alpha_T) I) (teacher-conditioned training start) on held-out inputs, reporting PSNR/SSIM over at least three seeds. If the extractor-only row matches full OSMamba within error bars, or if pure-noise starts lower PSNR by more than 0.1 dB versus teacher-conditioned starts, the DDPG's generative prior is not the source of the reported gain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The SOTA claim rests on the DDPG prior: Table 4(a) shows that removing the prior drops LCDP PSNR from 24.53 to 23.72, below CoTF (23.89), so without this module the headline result fails on LCDP. The only ablation isolating the generative diffusion is Table 4 settings (b) vs (c), and both rows use the spatial-domain SDPE/SDPG; there is no dual-domain extractor-only control (\"DDPE* without denoising\"), so the 0.52 dB gain attributed to the diffusion model is entangled with the spatial-to-dual-domain change. The mechanism also has an unaddressed train/inference shift: training starts the reverse chain at N(sqrt(bar-alpha_T) Z, (1 - bar-alpha_T) I), while inference starts at N(0, I). With T=4 and bar-alpha_T=0.01, the training start contains a 0.1Z signal component; unless Z is variance-normalized (not stated), the DDPG at inference is off-distribution. Because the UNet is fine-tuned to the distilled prior, a degraded prior directly reduces corrected-image quality. All ablations are reported on LCDP only, with no error bars or multi-seed runs, so the transfer of the DDPG benefit to MSEC and SICE is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes OSMamba, an exposure correction network that combines a frequency-domain Mamba with an omnidirectional spectral scanning mechanism (OS-Scan) and a dual-domain prior generator (DDPG). The model is trained in two stages: first, a teacher prior extractor (DDPE) uses ground-truth images to create compact priors that modulate the UNet; second, a student generator (DDPG) distills this teacher into a GT-free conditional diffusion model. Experiments on MSEC, SICE, and LCDP report state-of-the-art PSNR/SSIM, with ablations on LCDP supporting the contributions of OS-SSM and the prior generator.","tokens_in":16375,"tokens_out":8359,"duration_ms":78490,"significance":"If the results hold, OSMamba is a meaningful empirical advance: it is the first exposure correction network to apply Mamba with four-directional scanning in the Fourier domain, and the two-stage distillation of a GT-dependent prior into a GT-free diffusion generator is a clever way to inject generative detail restoration without inference-time ground truth. The reported gains over strong baselines such as CoTF are consistent across three datasets and the ablations show that both main components contribute. However, the evidence is weakened by a confounded ablation for the diffusion model, a train/inference distribution mismatch in the diffusion start, missing comparisons to a recent Mamba-based method, and the absence of error bars on small margins.","major_comments":[{"comment":"The ablation isolating the contribution of the generative diffusion model is confounded. The text states that setting (b) 'removes the latent diffusion model' and attributes the 0.52 dB drop to that removal, but setting (b) uses the spatial-domain extractor SDPE, not the dual-domain DDPE. Therefore the 0.52 dB difference between OSMamba (DDPE+DDPG) and setting (b) conflates two changes: the extractor domain (spatial vs dual) and the presence of the diffusion generator. A proper control, e.g., DDPE without the diffusion generator, is needed to attribute the gain specifically to the diffusion process. This is load-bearing because the second contribution claim rests on this ablation.","section":"Section 4.3, Table 4"},{"comment":"There is a train/inference mismatch in the diffusion start. Training initializes the reverse chain as N(sqrt(bar_alpha_T) Z, (1 - bar_alpha_T) I) with bar_alpha_T at most 0.01, so the start contains a non-negligible signal component from the ground-truth-based prior Z. Inference initializes from pure noise N(0, I), as stated in Section 3.5. With only T=4 steps and a two-layer denoiser, the model never sees a pure-noise starting point during training, and the paper does not analyze whether this distribution shift degrades the generated prior. This directly affects the validity of the DDPG at inference, which is critical for the SOTA claim on LCDP where removing the prior drops performance below CoTF.","section":"Section 3.5, Algorithm 1, and Section 4.1"},{"comment":"The paper does not compare quantitatively against ECMamba [15], a NeurIPS 2024 Mamba-based exposure correction method, even though the manuscript is a May 2025 version and cites this work in the related literature. Since the paper's central novelty is a Mamba variant, the SOTA claim requires a direct comparison to this concurrent method on at least the shared datasets (MSEC and SICE). Without it, the claim of state-of-the-art performance is incomplete.","section":"Table 1 and Related Work"},{"comment":"All ablations for the two main contributions are performed only on the LCDP dataset. The paper claims SOTA on MSEC and SICE, and the DDPG is a key component, but its benefit is not demonstrated on those datasets. Given that the margins over baselines vary across datasets, the authors should report at least one ablation (e.g., removing the prior or disabling the diffusion generator) on MSEC or SICE to support the generality of the contributions.","section":"Section 4.3"},{"comment":"All reported metrics are from single runs with no error bars, confidence intervals, or significance tests. Several improvements are small (e.g., 0.18 dB in Table 3, 0.21 dB in Table 4). Without an estimate of run-to-run variance, it is impossible to judge whether these gains are reliable. At minimum, the authors should report results over multiple seeds (e.g., three) for the ablations and the final model, or provide a statistical test for the main comparisons.","section":"Tables 1-4"}],"minor_comments":[{"comment":"The text says the baselines keep 'a comparable number of parameters,' but no parameter counts are given. Please report the actual parameter numbers for the OS-SSM, VSSM, and Attention variants to support this claim.","section":"Section 4.3, Table 2"},{"comment":"The notation for the amplitude and phase features is inconsistent: the text uses Ah(X), Ph(X) for inputs and \\tilde Ah(X), \\tilde Ph(X) for outputs, but the equation later refers to these as A_h(X) and P_h(X). Please standardize the notation.","section":"Section 3.3, Eq. (5)"},{"comment":"The caption contains a typo: 'Spatial Featurae Refinement' should be 'Spatial Feature Refinement.'","section":"Figure 3 caption"},{"comment":"The claim of a 'significant improvement of 1.01 dB' on SICE should specify the baseline (the table shows the gain over MMHT, which is the second-best average PSNR).","section":"Section 4.2"},{"comment":"The DDPE* is described as 'same structure as DDPE except for the input convolution,' but DDPE takes two concatenated images (Ierror and Igt) while DDPE* takes one. Please clarify the exact input-channel adjustment.","section":"Section 3.5"},{"comment":"The paper reports 7.5M parameters but does not provide FLOPs or inference time. Since the method combines SSMs and diffusion, reporting computational cost would help practitioners assess its practicality.","section":"Section 4.1"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a well-engineered architecture with a novel combination of spectral Mamba and diffusion-based prior distillation. The core ideas are plausible and the experimental results are promising, but the ablation that is supposed to isolate the diffusion contribution is confounded, and the train/inference mismatch in the diffusion generator needs explicit analysis. I also consider the missing comparison to ECMamba a notable gap for a 2025 version. The requested revisions (additional ablations, a control without the diffusion generator, and a discussion of the distribution shift) are feasible within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a competent, useful engineering paper. The new combination is omnidirectional spectral scanning (row, column, and both diagonal zig-zags in the Fourier amplitude/phase) and a dual-domain prior generator that distills a ground-truth-dependent prior into a GT-free diffusion prior. Each ingredient is borrowed, but the specific architecture is not in prior work, and the paper shows consistent gains across three benchmarks. The ablations are thoughtful: replacing OS-SSM with VSSM or attention drops PSNR, and removing the prior entirely costs 0.81 dB on LCDP. That is real evidence that the components do something.\n\nWhere the paper gets soft: the diffusion prior ablation is reported on LCDP only, with no error bars or multi-seed runs. More concretely, the paper claims removing the latent diffusion model causes a 0.52 dB drop by comparing setting (b) to full OSMamba, but that comparison also switches from spatial-domain to dual-domain extraction. The clean diffusion-only comparison (b vs c) is 0.31 dB, and the domain switch (c vs full) is 0.21 dB. That is a misreading of the table, not a fatal flaw, but it should be corrected. The train/inference diffusion-start mismatch (training begins with a tiny signal component, inference from pure noise) is not discussed; with T=4 and a small alpha_bar_T the impact is probably minor, but it deserves a sentence. The attention module inside OS-SSB is under-specified, and no code is released.\n\nThese are solid, fixable issues. The central SOTA claim is plausible; the experiments are internally consistent and the ablations broadly support the design choices. The citation pattern looks fair, with prior Mamba variants and diffusion-prior methods clearly acknowledged. This paper is for people working on exposure correction or applying Mamba to frequency-domain restoration; they will find it a useful baseline. It deserves a serious referee, not a desk rejection, but I would make acceptance conditional on code release, correcting the ablation attribution, and adding error bars or at least multi-seed runs on LCDP.","headline":"Solid subfield-level engineering: Mamba in the Fourier domain plus a distilled diffusion prior buys 0.2–1.4 dB on exposure correction, but the diffusion-ablations misattribute part of the gain.","tokens_in":16905,"tokens_out":2204,"would_cite":true,"duration_ms":41230,"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":"OSMamba claims state-of-the-art exposure correction by pairing four-direction spectral Mamba scanning with a distilled dual-domain diffusion prior.","keywords":["exposure correction","state space models","Mamba","frequency domain","diffusion prior","image restoration","dual-domain prior","spectral scanning"],"falsifier":"Take the hardest quartile of under- and over-exposed test images, replace the DDPG prior at inference with a constant vector of the same shape, and measure whether PSNR drops by substantially more than the 0.52 dB ablation gap; a negligible drop would mean the distilled prior is not what restores the lost details.","tokens_in":15856,"feed_emoji":"📷","tokens_out":7873,"duration_ms":69373,"temperature":0.7,"pith_summary":"Exposure correction in extreme under- or over-exposed images is hard because local convolutions cannot capture long-range dependencies in the frequency spectrum and regression losses cannot recreate lost high-frequency detail. The paper argues that both problems can be fixed together: a state space model scanning the amplitude and phase of deep features along rows, columns, and diagonals captures global spectral structure, and a compact latent diffusion prior, distilled from a ground-truth-dependent extractor into a ground-truth-free generator, supplies the missing detail. The resulting network, OSMamba, is reported to give the highest PSNR and SSIM on the MSEC, SICE, and LCDP benchmarks, with a 1.01 dB gain over the second-best method on SICE. A sympathetic reader should take from this that frequency-domain global modeling plus generative detail priors is a promising direction for exposure correction.","feed_headline":"New network leads three exposure-correction benchmarks","feed_subtitle":"It scans Fourier amplitude and phase in four directions and injects a distilled diffusion prior to restore lost detail.","key_machinery":"The key machinery is the Omnidirectional Spectral State Space Model (OS-SSM) together with the Dual-Domain Prior Generator (DDPG). OS-SSM takes the 2D FFT of deep features, keeps half the spectrum, and runs two parallel Mamba modules on the amplitude and phase with OS-Scan, four continuous symmetric scanning directions (row, column, positive diagonal, negative diagonal) that respect the spectrum's symmetry and continuity, before recombining with the inverse FFT and gating. DDPG is a conditional latent diffusion model that starts from random Gaussian noise and performs T=4 denoising steps conditioned on a dual-domain feature extracted from the input image alone, producing a prior vector that is injected into every OS-SSM via affine transformation; in training it distills the ground-truth-dependent DDPE prior. These two components carry the argument: the first provides global receptive fields in the frequency domain for illumination and structure correction, and the second supplies the external, degradation-free information needed to reconstruct details lost in extreme exposure.","core_discovery":"The central discovery claimed is that an exposure correction network built on an Omnidirectional Spectral State Space Model (OS-SSM) and a Dual-Domain Prior Generator (DDPG) achieves state-of-the-art results on multiple- and mixed-exposure benchmarks. OS-SSM treats the half-spectrum of deep features as two parallel streams, amplitude (lighting and color) and phase (structure), and applies Mamba's S6 selective scan along four symmetric trajectories: rows, columns, positive diagonals, and negative diagonals, exploiting the central symmetry of the Fourier spectrum. DDPG is a conditional diffusion model with only T=4 denoising steps that produces a compact prior vector from the input image alone; it is trained by distilling the prior extracted by the ground-truth-dependent DDPE teacher, and each OS-SSM applies the prior as an affine modulation. The paper reports best PSNR and SSIM of 23.78 and 0.8767 on MSEC, 22.82 and 0.7201 on SICE, and 24.53 and 0.8773 on LCDP, with visual comparisons showing restored detail and natural color in severely degraded regions.","pith_inferences":["The paper's own ablation (Table 4) attributes 0.52 dB of the LCDP gain to the latent diffusion model, but only on one dataset and without error bars; a reasonable next step is to test whether that gap is stable across random seeds and input severity levels.","Since the prior is only 1x1xM, it may be that most of the correction signal carried by the prior is global, such as global illumination and color statistics, rather than spatially localized detail; a spatial breakdown of the prior's effect would settle this.","The same distillation-from-a-dual-domain-teacher recipe could be applied to other ill-posed restoration problems such as deblurring or deraining, where high-frequency detail is also missing and could be supplied by a compact generative prior.","For deployment, the T=4 denoising steps mean the generative prior adds little latency; an interesting extension would be to measure how prior quality changes as T is reduced to 1 or 2, which would reveal how much of the diffusion process is genuinely load-bearing."],"forward_implications":["If the reported results hold, OSMamba becomes the reference point that future exposure correction methods must beat on MSEC, SICE, and LCDP.","The four-direction spectral scan suggests that diagonal dependencies in the Fourier spectrum carry information that row- and column-only scans miss, which may transfer to other restoration tasks that decouple amplitude and phase.","The two-stage teacher-to-student distillation shows a path for turning priors that need ground truth into deployable, input-only priors while keeping the detail-restoration benefit.","Because Mamba scanning is linear in sequence length, the frequency-domain global modeling comes without the quadratic cost of attention, which matters for high-resolution images."],"supporting_citations":[{"why":"Supplies the selective state-space (S6) layer that OS-SSM adapts to the frequency domain.","marker":"[19]"},{"why":"Provides the structured state-space formulation and zero-order hold discretization used in Equation (2).","marker":"[20]"},{"why":"Defines the Cross-Scan mechanism that OSMamba extends, and is the baseline in the scanning ablation.","marker":"[43]"},{"why":"The frequency-domain predecessor whose amplitude/phase decomposition and local-convolution limitation OSMamba addresses.","marker":"[29]"},{"why":"Provides the noising schedule and reverse denoising formulation used by the dual-domain prior generator.","marker":"[26]"},{"why":"The diffusion-prior approach that DDPG's distillation design draws on.","marker":"[61]"},{"why":"The latent-diffusion restoration method used as comparison for the dual-domain prior design.","marker":"[24]"},{"why":"Supplies the VSSM block and the cross-scan baseline replaced in the OS-SSB ablation.","marker":"[22]"},{"why":"A strong recent baseline whose LCDP PSNR OSMamba exceeds by 0.64 dB.","marker":"[38]"}],"fun_headline_variants":["Mamba meets Fourier: new exposure correction SOTA","Omnidirectional spectral Mamba fixes exposure","Dual-domain prior boosts exposure correction","Amplitude and phase scanned: OSMamba tops exposure benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the distilled dual-domain prior generator, which learns from priors extracted with access to the ground-truth image, produces a prior from the input image alone that is as informative as the teacher prior at test time.","fun_headline_variants_meta":{"raw":{"variants":["Mamba meets Fourier: new exposure correction SOTA","Omnidirectional spectral Mamba fixes exposure","Dual-domain prior boosts exposure correction","Amplitude and phase scanned: OSMamba tops exposure benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000516,"raw_usage":{"total_tokens":2525,"prompt_tokens":985,"completion_tokens":1540,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":1479}},"tokens_in":601,"tokens_out":1540,"duration_ms":12002,"temperature":1.0,"reasoning_tokens":1479,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:52:18.421863+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the hardest quartile of under- and over-exposed test images, replace the DDPG prior at inference with a constant vector of the same shape, and measure whether PSNR drops by substantially more than the 0.52 dB ablation gap; a negligible drop would mean the distilled prior is not what restores the lost details.","supporting_citations":[{"cited_title":"Vmamba: Visual state space model","cited_arxiv_id":null,"evidence_quote":"Defines the Cross-Scan mechanism that OSMamba extends, and is the baseline in the scanning ablation."},{"cited_title":"Deep fourier-based exposure correction network with spatial- frequency interaction","cited_arxiv_id":null,"evidence_quote":"The frequency-domain predecessor whose amplitude/phase decomposition and local-convolution limitation OSMamba addresses."},{"cited_title":"Diffir: Efficient diffusion model for image restoration","cited_arxiv_id":null,"evidence_quote":"The diffusion-prior approach that DDPG's distillation design draws on."},{"cited_title":"Mambair: A simple baseline for im- age restoration with state-space model","cited_arxiv_id":null,"evidence_quote":"Supplies the VSSM block and the cross-scan baseline replaced in the OS-SSB ablation."},{"cited_title":"Real-time expo- sure correction via collaborative transformations and adap- tive sampling","cited_arxiv_id":null,"evidence_quote":"A strong recent baseline whose LCDP PSNR OSMamba exceeds by 0.64 dB."}],"review_version":1}