{"id":"9d5e6c81-5aa2-44bb-919e-0ae7550f2f98","arxiv_id":"2507.01838","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A 4K-parameter reparameterized CNN with square-transform features, dual-path attention, and a variance-weighted loss reaches about 1,100 FPS on image enhancement benchmarks.","lead":"MobileIE is a four-thousand-parameter convolutional network that enhances low-light, underwater, and raw camera images in real time, claiming up to 1,100 frames per second on a consumer GPU. It shows that extremely small models can stay competitive with much larger enhancement networks, though code and measurement details are not yet available.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central speed/quality claim depends on unverified dataset splits and missing artifacts; Section 4.1 does not state train/test partitions for LOLv1/LOLv2, and no code/checkpoints exist, so the reported PSNR/FPS trade-off is not yet reproducible.","rationale":"The paper's central claim is an empirical Pareto claim: at ~4K parameters and ~0.9 ms/1120 FPS, MobileIE matches or beats prior lightweight IE methods in quality across three tasks. For this to hold, the reported PSNR/SSIM/FPS numbers must be measured under a clean protocol. The reader's weakest assumption points to this protocol; I agree that it is the core risk. Section 4.1 is unusually terse: it names datasets without any explicit split indices, and the supplement that would contain the channel configuration is absent from the arXiv version. Without code or checkpoints, there is no way to check that the loss implemented is the one described—a non-trivial issue because Eq. (8) defines a global mean/variance while the text calls it 'local.' This is not a refutation: the architecture is simple, re-parameterization is a known technique, and the numbers are internally plausible given SYELLE's 1059 FPS with 5K params. But the marginal speed gain over SYELLE (~6%) is small, so even a minor protocol difference (e.g., using a different input resolution or averaging multiple runs) could erase the claimed advantage. The reader's CONDITIONAL verdict is therefore the right call. My concrete test would settle the concern by checking whether the standard-split reproduction matches the table; until then, the central claim remains unverified, not disproven. I give the paper credit for the honest detail in Table 7 (train vs inference cost) and the direct comparison to SYELLE, but those do not substitute for artifacts.","tokens_in":17650,"tokens_out":7624,"duration_ms":82456,"concrete_test":"After code/checkpoints are released, run the official model on the standard LOLv1 test split (15 images) and the LOLv2-Real test split (70 images) with batch size 1 on an RTX 4090, and compare PSNR/SSIM/FPS to Table 1. Independently, re-train MobileIE on the standard training splits and verify that the reported numbers are within 0.5 dB PSNR and 10% FPS; also re-implement Eq. (8) with a genuine local window (e.g., 7×7) and as written (global) to see which matches Table 4's LVW row. If the standard-split results deviate beyond these tolerances, the headline trade-off is not verified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Claim: MobileIE achieves the best speed/quality trade-off across LLE, UIE, and ISP with ~4K parameters (Tables 1–3). The load-bearing condition is that the benchmark numbers are measured on standard splits and with a faithful implementation. Section 4.1 only says 'LOLv1 and LOLv2 datasets are used for both training and testing'; it does not specify the train/test split (e.g., LOLv1 485/15, LOLv2-Real 619/70), nor whether LOLv2-Synthetic is included. If any test image was seen during training, the reported PSNR/SSIM/LPIPS gains (e.g., 23.62 dB on LOLv1, 25.08 on LOLv2-Real) are inflated and the trade-off claim collapses. The missing appendix containing exact channel configurations and the absence of released code/checkpoints prevent independent verification of the 0.895 ms/1120 FPS GPU numbers. An internal inconsistency strengthens this concern: the 'Local Variance Weighted' loss in Eq. (8) computes a single global mean/variance over the entire H×W map, not per-pixel local statistics, contradicting the text and the 'local' name; a re-implementation following the equations would use a different loss than the one described. Since the paper provides no executable artifacts, the central claim is currently unfalsifiable.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MobileIE, a CNN with roughly 4K parameters for image enhancement, designed for real-time use on mobile devices. The architecture combines multi-branch reparameterized convolutions (MBRConv), an Incremental Weight Optimization (IWO) training schedule, a Feature Self-Transform (FST) block, a Hierarchical Dual-Path Attention (HDPA) module, and a Local Variance-Weighted (LVW) loss. Experiments are reported on low-light enhancement (LOLv1, LOLv2-Real), underwater enhancement (UIEB), and raw-to-RGB ISP (ZRR), with the claim that MobileIE achieves up to about 1,100 FPS on a GPU, around 6.7 ms on a Snapdragon 8 Gen 3 SoC, and competitive or better PSNR/SSIM/LPIPS than existing lightweight methods.","tokens_in":17900,"tokens_out":6461,"duration_ms":70810,"significance":"If the results are reproducible, the contribution is practically useful: a roughly 4K-parameter model that runs in under a millisecond on a GPU and in real time on a mobile SoC while matching or exceeding the quality of much larger lightweight enhancement networks would be a meaningful engineering advance. The paper also includes per-module ablations and visualizations, and the proposed modules are simple and hardware-friendly. However, the central speed/quality claim is currently conditional on dataset split details, benchmarking methodology, and code/checkpoint availability, none of which are provided in sufficient detail. The significance is therefore contingent on closing these reproducibility gaps.","major_comments":[{"comment":"The dataset protocol is not specified. The text says \"LOLv1 [76] and LOLv2 [82] datasets are used for both training and testing,\" but it does not state the train/test partition (e.g., 485/15 for LOLv1, 619/70 for LOLv2-Real), whether LOLv2 includes the Synthetic subset, or the image resolution and preprocessing for each dataset. The same omission applies to UIEB and ZRR. Without this information, the reported PSNR/SSIM values (e.g., 23.62 dB on LOLv1 and 25.08 on LOLv2-Real in Table 1) cannot be reproduced, and a split that leaks test images into training would invalidate the headline trade-off claim. Please specify exact splits, random seeds, and preprocessing steps.","section":"4.1"},{"comment":"Equation (8) does not match the \"local\" description in the text. The formulas for µ_{m,n} and σ²_{m,n} sum over all H×W spatial positions, so µ and σ are single global scalars per image/channel, not per-pixel local statistics. The subscripts are misleading, and the claim that the loss \"adapts to local variability\" is not supported by the equation. If the implementation follows the equation, the loss is global-mean/variance weighted; if a local window is used, Eq. (8) must be rewritten. Since Table 4 attributes a performance gain to LVW loss, the correct form of the loss is load-bearing and must be clarified. In addition, the second line of Eq. (8) has a missing parenthesis: the sum of (Δ_{m,n} − µ_{m,n}) should be squared as a whole.","section":"3.5"},{"comment":"The manuscript refers to an Appendix for \"detailed architecture settings,\" but no appendix is present in the submitted version. Furthermore, the GPU (RTX 4090) and SoC (Snapdragon 8 Gen 3) latency numbers in Tables 1–3 are reported without describing the measurement methodology: batch size, warm-up iterations, inference framework, precision, number of runs, and whether the mobile measurements used the SoC's NPU, GPU, or CPU. No code or checkpoints are provided, and no standard deviations over repeated runs are reported. These omissions make the central claim of up to 1,100 FPS with ~4K parameters unfalsifiable. Please release the code and checkpoints, provide the appendix with exact channel configurations and layer-wise parameter/FLOP counts, and describe the benchmarking protocol in enough detail to be reproduced.","section":"3.1"},{"comment":"The IWO training schedule is underspecified. Section 4.1 lists optimizer, learning rate, warm-up, and total epochs, but does not state when IWO is activated, how W_pre is chosen and frozen in Eq. (1), or how W_learn is initialized. Section 4.3 later says W_pre is \"the well-performing weight obtained after the first 1000 epochs,\" which suggests a phase split that should have been in Section 4.1. The IWO phase epoch is a key hyperparameter of the proposed method, and the ablation results in Table 4 (MBRConv No IWO vs. Ours) depend on it. Please state the exact IWO protocol in the experimental settings.","section":"4.1"},{"comment":"Table 7 is headed \"RTX 3090,\" while Section 4.2 and Tables 1–3 state that latency tests were run on a single NVIDIA 4090. If the table and the earlier tables use different hardware, the numbers are not directly comparable; if it is a typo, it should be corrected. In addition, the text in Section 4.3(3) says HDPA \"achieved a PSNR gain of +3.08 dB,\" but Table 6 contains no baseline row without an attention module, so the gain cannot be read from the table. Please add the missing baseline and ensure hardware labels are consistent across all tables.","section":"4.3 (Table 7)"}],"minor_comments":[{"comment":"The claim of being \"the first\" to achieve real-time IE at up to 1,100 FPS should be qualified, because Table 1 lists SYELLE at 1,059.7 FPS, which is very close; the novelty should be framed around the combined speed-quality-parameter trade-off rather than a raw FPS record.","section":"Abstract, Section 1"},{"comment":"Section 4.2 refers to \"the ZRR [44] dataset\" for ISP experiments, but Table 3 and reference [36] identify ZRR as the dataset from PyNet; please correct the citation.","section":"4.2"},{"comment":"The first row of Table 4, labeled \"Only inference network,\" is not defined in the text; please state what training setup this row corresponds to (e.g., no loss term or random initialization).","section":"Table 4"},{"comment":"The figure label \"FivaA+\" contains a typo and should read \"FiveA+\".","section":"Figure 6"},{"comment":"In Eq. (5), W_g = F * A_g is a weighted feature map, not a set of attention weights; naming it \"W\" is confusing because A_l is later multiplied with A_g and F in Eq. (6). Please rename W_g to something like F_g to avoid suggesting it is a weight tensor.","section":"3.4"},{"comment":"The SCORE metric [38] is used throughout but never defined in the paper; given the large SCORE differences reported (e.g., 92.855 vs. 9.604 in Table 1), please include the formula or a brief explanation so readers can interpret these values.","section":"Tables 1–3, 6"},{"comment":"In Eq. (9), the constant ε is described only as \"a small constant\"; please give its value, and clarify whether the tanh normalization is applied per channel or across the whole image.","section":"3.5"}],"recommendation":"major_revision","confidential_remarks":"The core issue is reproducibility: the missing appendix, unspecified dataset splits, undocumented benchmarking setup, and the inconsistency between Eq. (8) and the 'local variance' description are all addressable in revision, but until they are fixed the claimed speed-quality trade-off cannot be independently checked. The authors should also position the contribution relative to SYELLE, which already reports ~1,060 FPS, so that the novelty is clear."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Peer,\n\nQuick take: MobileIE is a plausible small-step improvement over SYELLE (4K params vs 5K, ~1120 FPS vs 1059, better PSNR on LOLv1), built from known components. The ablation work is decent and the numbers are internally consistent. But the paper currently can't be checked: no code, no checkpoints, the LLE dataset split is unstated, and the LVW loss has a naming/formula mismatch. These are fixable, so I'd send it to referees, with a stern request for artifacts and protocol details.\n\nWhat's genuinely new is the combination and the specific measured trade-off. The architecture recycles RepVGG-style multi-branch reparam, StarNet's square function, and a pooling-based dual attention; IWO is just a continued-training schedule with a frozen prior. That's fine—incremental work can be useful. The experiments cover three tasks and the ablations actually test each module.\n\nSoft spots, in rough order of importance:\n\n1. No code or checkpoints. For a claims-driven efficiency paper, that's the main gap. Without it, the 1120 FPS number is an assertion.\n2. Section 4.1 says LOLv1 and LOLv2 are 'used for both training and testing' without giving the split. If they trained on the test images, the PSNR gains are inflated. Standard splits exist and should be cited.\n3. The LVW loss: Eq. (8) sums over the full H×W map, producing one mean/variance per sample and channel. That's global, not local. The text calls it 'local variance' and describes per-pixel statistics. A reader re-implementing from the equations will get a different loss than the one in prose.\n4. Latency measurements: no detail on how GPU/SoC timing was done. For a paper whose selling point is speed, that's a meaningful omission.\n5. The 'first to achieve 1,100 FPS' line doesn't hold up next to SYELLE's 1,059 FPS; that's a rhetorical overreach, not a technical flaw.\n\nAre these fatal? No. The internal evidence is self-consistent, the improvements over SYELLE are plausible, and the ablations don't show anything weird. The missing artifacts and split ambiguity are exactly what a good referee can force the authors to fix.\n\nWho would get value: anyone working on mobile or real-time image enhancement, especially those tracking the SYELLE line. The paper is a candidate for a systems-style venue, not a methods breakthrough.\n\nMy recommendation: accept for peer review, but with a clear message that code, checkpoints, and a precise evaluation protocol are prerequisites for acceptance. If the artifacts show up and the splits are standard, I'd take the numbers seriously.","headline":"Plausible incremental beat over SYELLE, but the missing artifacts and a sloppy loss definition make the headline numbers unverifiable until the authors release code and split details.","tokens_in":18484,"tokens_out":3770,"would_cite":false,"duration_ms":40726,"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 4K-parameter convolutional network performs real-time image enhancement at up to 1,100 FPS while matching the quality of models tens to hundreds of times larger.","keywords":["image enhancement","low-light enhancement","underwater image enhancement","image signal processing","reparameterization","mobile inference","lightweight CNN","real-time enhancement"],"falsifier":"Once code and checkpoints are made available, run the released MobileIE checkpoint on the standard LOLv1 test split and measure PSNR plus end-to-end latency on the same RTX 4090 and Snapdragon 8 Gen 3 devices used for the comparison methods; if the PSNR falls materially below the reported 23.62 dB or the 0.895 ms GPU latency cannot be reproduced, the central speed–quality trade-off claim fails.","tokens_in":17395,"feed_emoji":"⚡","tokens_out":6023,"duration_ms":61077,"temperature":0.7,"pith_summary":"MobileIE claims that a convolutional network with roughly 4,000 parameters can perform real-time image enhancement—low-light correction, underwater restoration, and raw-to-sRGB processing—at up to 1,100 frames per second on a GPU and about 6.7 ms on a Snapdragon 8 Gen 3 phone, while matching or exceeding the quality of lightweight models that are tens to hundreds of times larger. The paper argues that reparameterization, a frozen-plus-learnable weight update, a quadratic feature transform, a dual-path attention module, and a variance-weighted loss together let a tiny network escape the usual quality floor. If correct, this would move real-time enhancement onto ordinary mobile hardware and suggest that extreme parameter compression does not have to cost image quality.","feed_headline":"A 4K-parameter network enhances images at 1,100 FPS","feed_subtitle":"Matches or beats far larger lightweight models on low-light, underwater, and raw-to-sRGB tasks in under a millisecond.","key_machinery":"The pipeline is built on five components that work together. MBRConv is a multi-branch reparameterized convolution (branches of 3x3, 5x5, and 1x1 with parallel BatchNorm layers) that folds into a single standard convolution at inference, cutting parameters to about 4K with no added inference cost. Incremental Weight Optimization freezes a previously trained weight matrix and adds a learnable residual delta to it, letting the network keep improving late in training. Feature Self-Transform squares features channel-wise and applies a learnable scale and bias, adding nonlinear expressiveness beyond linear convolutions. Hierarchical Dual-Path Attention combines average-pooled global channel weights with max-pooled local weights and multiplies them onto the feature map. Local Variance-Weighted Loss reweights each pixel's L1 error by the hyperbolic tangent of its deviation from the local error mean divided by the local variance, reducing the influence of outlier pixels. These mechanisms carry the paper's claim that extreme compactness does not force a quality penalty.","core_discovery":"On the paper's terms, the central discovery is that a reparameterized CNN with 4,047–4,104 parameters achieves the best speed–quality trade-off reported so far across three image enhancement tasks. On the LOLv1 and LOLv2-Real low-light benchmarks it reaches 23.62 dB and 25.08 dB PSNR, respectively, with 0.895 ms GPU latency and 1,120 FPS; on UIEB it reaches 22.81 dB PSNR at 0.910 ms, and on ZRR it reaches 21.43 dB PSNR at 1.020 ms, all while running at roughly 6.7–14.4 ms on a Snapdragon 8 Gen 3 phone. The authors attribute this to MBRConv folding multi-branch training structure into a single inference convolution, the Incremental Weight Optimization strategy that keeps a frozen prior weight and learns a residual delta, the Feature Self-Transform's quadratic interaction, the Hierarchical Dual-Path Attention fusing global and local statistics, and the Local Variance-Weighted loss that down-weights outliers. They claim this is the first demonstration of real-time image enhancement inference at up to 1,100 FPS with competitive quality.","pith_inferences":["If the central claim holds, the practical bottleneck for mobile image enhancement shifts from model size to data and training protocol: at 4K parameters the architecture is near the floor, so further gains would have to come from better losses, augmentation, or task-specific priors.","The Feature Self-Transform's squaring operation is a parameter-free nonlinearity that could transfer to other lightweight low-level vision tasks, such as super-resolution or deblurring, where fine detail preservation matters.","The Incremental Weight Optimization recipe—freeze a warm-started weight and learn a residual delta—could be a general cure for late-training stagnation in other compact networks, though the paper only demonstrates it inside MBRConv.","Because the headline SCORE advantage is dominated by parameter count and latency, a benchmark that weights quality more heavily could reorder the ranking; that is a consequence of the metric, not necessarily a weakness of the model."],"forward_implications":["A single 4K-parameter model can run low-light enhancement, underwater restoration, and raw-to-sRGB ISP in about 0.9 ms on a GPU and 6.7–14.4 ms on a Snapdragon 8 Gen 3 phone, making real-time on-device enhancement practical.","MobileIE matches or exceeds the PSNR and SSIM of prior lightweight models such as IAT, FiveA+, and SYELLE/SYEISP while using a fraction of their parameters and latency.","The reparameterization and cost-free training strategies (IWO, LVW) mean the training-time machinery adds no inference overhead, so the deployed model is a plain small CNN.","The same architecture, without task-specific changes, handles three different degradation types, supporting the idea of a universal lightweight enhancer.","The reported speeds open the door to real-time enhancement of UHD (2K–8K) images on mobile platforms, which the paper identifies as a target."],"supporting_citations":[{"why":"The prior real-time multi-task low-level vision network that MobileIE directly benchmarks against and extends.","marker":"[28]"},{"why":"The lightweight transformer for low-light enhancement that serves as the main parameter and speed comparison anchor.","marker":"[20]"},{"why":"The 9K-parameter underwater enhancement method that MobileIE surpasses with 45% of its parameters.","marker":"[41]"},{"why":"The RepVGG reparameterization approach that MBRConv adapts for image enhancement.","marker":"[24]"},{"why":"The ECBSR reparameterized block with BatchNorm for low-level vision, used as an ablation comparison.","marker":"[92]"},{"why":"The star-operation source for the quadratic interaction used in the Feature Self-Transform module.","marker":"[52]"},{"why":"The SCORE metric that defines the efficiency-performance trade-off used in the headline comparisons.","marker":"[38]"},{"why":"The UIEB underwater benchmark dataset and protocol for the underwater enhancement evaluation.","marker":"[44]"},{"why":"The LOLv1 low-light benchmark dataset and protocol for the low-light enhancement evaluation.","marker":"[76]"},{"why":"The PyNet mobile ISP baseline and ZRR dataset used for the raw-to-sRGB evaluation.","marker":"[36]"}],"fun_headline_variants":["4K-param CNN hits 1,100 FPS for mobile image enhancement","Ultra-light 4K-param net redefines mobile image enhancement speed","Tiny CNN achieves best speed-quality trade-off in image enhancement","First 1,100 FPS image enhancement on mobile with 4K-param CNN","MobileIE: 4K params, 1,100 FPS, best speed-quality trade-off"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported speed and quality numbers depend on the evaluation being run under fair, identical conditions on standard train/test splits; the paper does not yet provide code, checkpoints, or measurement details, so a failure of those conditions would change the headline PSNR and FPS figures.","fun_headline_variants_meta":{"raw":{"variants":["4K-param CNN hits 1,100 FPS for mobile image enhancement","Ultra-light 4K-param net redefines mobile image enhancement speed","Tiny CNN achieves best speed-quality trade-off in image enhancement","First 1,100 FPS image enhancement on mobile with 4K-param CNN","MobileIE: 4K params, 1,100 FPS, best speed-quality trade-off"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000919,"raw_usage":{"total_tokens":3963,"prompt_tokens":982,"completion_tokens":2981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":598,"completion_tokens_details":{"reasoning_tokens":2873}},"tokens_in":598,"tokens_out":2981,"duration_ms":24768,"temperature":1.0,"reasoning_tokens":2873,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T20:42:01.399070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Once code and checkpoints are made available, run the released MobileIE checkpoint on the standard LOLv1 test split and measure PSNR plus end-to-end latency on the same RTX 4090 and Snapdragon 8 Gen 3 devices used for the comparison methods; if the PSNR falls materially below the reported 23.62 dB or the 0.895 ms GPU latency cannot be reproduced, the central speed–quality trade-off claim fails.","supporting_citations":[{"cited_title":"Syenet: A simple yet effective net- work for multiple low-level vision tasks with real-time per- formance on mobile device","cited_arxiv_id":null,"evidence_quote":"The prior real-time multi-task low-level vision network that MobileIE directly benchmarks against and extends."},{"cited_title":"Five a+ net- work: You only need 9k parameters for underwater image enhancement","cited_arxiv_id":null,"evidence_quote":"The 9K-parameter underwater enhancement method that MobileIE surpasses with 45% of its parameters."},{"cited_title":"Edge-oriented convolution block for real-time super resolution on mobile devices","cited_arxiv_id":null,"evidence_quote":"The ECBSR reparameterized block with BatchNorm for low-level vision, used as an ablation comparison."},{"cited_title":"Rewrite the stars","cited_arxiv_id":null,"evidence_quote":"The star-operation source for the quadratic interaction used in the Feature Self-Transform module."},{"cited_title":"Learned smartphone isp on mobile gpus with deep learning, mobile ai & aim 2022 challenge: report","cited_arxiv_id":null,"evidence_quote":"The SCORE metric that defines the efficiency-performance trade-off used in the headline comparisons."},{"cited_title":"An underwater image enhancement benchmark dataset and beyond","cited_arxiv_id":null,"evidence_quote":"The UIEB underwater benchmark dataset and protocol for the underwater enhancement evaluation."},{"cited_title":"Deep retinex decomposition for low-light enhancement","cited_arxiv_id":null,"evidence_quote":"The LOLv1 low-light benchmark dataset and protocol for the low-light enhancement evaluation."},{"cited_title":"Replac- ing mobile camera isp with a single deep learning model","cited_arxiv_id":null,"evidence_quote":"The PyNet mobile ISP baseline and ZRR dataset used for the raw-to-sRGB evaluation."}],"review_version":1}