{"id":"ac2a1cb9-2568-41f8-b5a5-998613308d7b","arxiv_id":"2504.15649","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"RepNet-VSR reports 27.79 dB PSNR on REDS for 4x video super-resolution in 103 ms per 10 frames on a MediaTek NPU, but the reported scores do not match the paper's stated scoring formula.","lead":"The authors propose RepNet-VSR, a video super-resolution model for mobile NPUs that combines structural reparameterization with neural architecture search. The paper reports a better quality-speed trade-off than previous winners of the MAI video super-resolution challenge, but key runtime measurements and score calculations are internally inconsistent.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Composite scores in Tables 2–3 do not follow the paper's own Eq. (1), so the claimed superiority over EVSRNet/RCBSR lacks a reproducible numerical basis.","rationale":"The reader's REJECT verdict is supported, but I locate the most load-bearing concern slightly differently. The reader's stated weakest assumption is that the 103 ms runtime was measured under the official MAI2025 protocol on the MediaTek NPU; my analysis instead targets the direct arithmetic contradiction between the paper's own scoring formula (Eq. 1) and the Score values in Tables 2 and 3. This internal inconsistency is decisive because the headline comparison reduces to those Score numbers, and it does not depend on resolving the hardware provenance in Section 4.3 versus Section 4.5. The reader does mention the scoring-formula inconsistency in the rationale, so there is partial agreement. Given that the central quantitative claim is not reproducible from the paper's equations, the REJECT verdict stands without adjustment. The technical building blocks (NAS, structural reparameterization, 1×1 convolution before depth-to-space) are not the issue; the issue is that the paper's reported results do not follow its own stated metric.","tokens_in":7318,"tokens_out":7212,"duration_ms":63603,"concrete_test":"Recompute every Score entry in Tables 2 and 3 from the paper's own Eq. (1), using the PSNR and runtime values in the same row. If any entry differs from 22·(PSNR−27)/runtime by more than rounding (say 0.1% relative), then either Eq. (1) or the Score column is misreported. In that case, obtain the official MAI2025 scoring formula and the raw AI Benchmark logs from the authors or organizers to determine which quantity (formula, score, or runtime) is wrong, and re-run the comparison before accepting the champion-beating claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that RepNet-VSR beats previous MAI champions on composite score. That comparison rests on the Score column of Table 3 (0.029 vs 0.0174/0.0154). The paper states in Section 3.1 that the challenge score is Score = 22·(PSNR−27)/runtime. Substituting the paper's own Table 3 values gives EVSRNet 22×0.42/103 ≈ 0.0897 (reported 0.0174), RCBSR 22×0.28/95.8 ≈ 0.0643 (reported 0.0154), and RepNet-VSR 22×0.79/103 ≈ 0.1687 (reported 0.029). Table 2 has the same problem (e.g., row 2: 22×0.79/89.6 ≈ 0.194 vs 0.0334). The discrepancies are not a constant factor, so no simple unit conversion rescues the tables. Either Eq. (1) is not the scoring formula actually used, or the Score column is wrong; in both cases the claimed '66.7%/88.3% higher' comparisons are not reproducible from the paper's data. A secondary contradiction: Section 4.3 reports runtime evaluation on a Qualcomm Snapdragon 870 GPU because the MediaTek device was unavailable, while Section 4.5 and the abstract report 103 ms on the MediaTek NPU with no measurement procedure. This further undermines the runtime input to the score, but the arithmetic inconsistency alone is sufficient to invalidate the central claim as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes RepNet-VSR, a reparameterizable convolutional architecture for 4x video super-resolution on mobile NPUs. The architecture uses NAS to select the number of channels and RepBlocks, 1x1 convolutions for channel compression before depth-to-space, and multi-level feature fusion via channel concatenation. On the REDS validation set the authors report 27.79 dB PSNR and 103 ms per 10 frames on a MediaTek Dimensity NPU, and a MAI2025 composite score of 0.029 that is claimed to be 66.7% higher than EVSRNet and 88.3% higher than RCBSR. The paper also presents an ablation over channel/block configurations and a qualitative visual comparison.","tokens_in":7675,"tokens_out":7724,"duration_ms":69427,"significance":"If substantiated, the method would be a useful contribution to efficient edge VSR: the use of reparameterization, NAS-guided configuration, and 1x1 channel compression before upsampling is a sensible design direction, and the reported PSNR/runtime combination would be competitive. The paper is commendable for including multiple model configurations in Table 2 and for making the architectural recipe in Figures 1 and 2 concrete. However, the central comparative claim depends entirely on score and runtime values that cannot be reproduced from the paper's own formulas and measurement statements; until those are corrected and verified on the official hardware, the significance of the claimed superiority is not established.","major_comments":[{"comment":"The composite scores in Tables 2 and 3 do not follow the scoring formula stated in Eq. (1). For example, substituting EVSRNet's reported PSNR 27.42 and NPU runtime 103 ms into Eq. (1) gives 22*0.42/103 ≈ 0.0897, not the reported 0.0174; RCBSR gives 22*0.28/95.8 ≈ 0.0643, not 0.0154; and the proposed model gives 22*0.79/103 ≈ 0.169, not 0.029. Table 2 has the same discrepancy, e.g. row 2 gives 22*0.79/89.6 ≈ 0.194 versus the reported 0.0334. The discrepancies are not a constant rescaling, so no simple unit conversion repairs them. Consequently, the claimed 66.7% and 88.3% improvements over the previous champions are not reproducible from the paper's own data; either Eq. (1) is not the scoring formula actually used, or the Score columns are incorrect, and in both cases the central comparative claim is not supported.","section":"Section 3.1, Eq. (1), Tables 2-3"},{"comment":"The runtime measurement that feeds the score is not supported. Section 4.3 explicitly states that the MediaTek hardware was unavailable and that efficiency assessments were conducted on a Qualcomm Snapdragon 870 GPU, yet the abstract and Section 4.5 report 103 ms per 10 frames on a MediaTek Dimensity NPU without describing any measurement procedure. Section 4.4 labels Table 2 as Snapdragon 870 GPU results, while Table 3 is labeled MediaTek NPU, and no protocol for the NPU benchmark (delegate, quantization, number of runs, warm-up) is given. Moreover, the architecture selection in Table 2 uses runtimes from a different hardware platform than the final NPU comparison in Table 3, so the selected 'optimal' configuration is not validated on the target hardware. This undermines the runtime input to Eq. (1) and the comparability of the claimed NPU results.","section":"Section 4.3 vs. Section 4.5"}],"minor_comments":[{"comment":"The search space lists 'The number of channels 0,1,..,32'; a channel count of 0 is degenerate and presumably a typo for a positive range.","section":"Table 1"},{"comment":"Equation (2) uses symbols θ and ψ without defining them, and the value of the regularizer weight λ is not reported; please clarify the notation and provide the loss configuration used in the NAS phase.","section":"Equation (2)"},{"comment":"The heading contains a typo ('Comparision' should be 'Comparison'), and Figure 3 would benefit from explicit crop locations or zoomed insets to support the claimed qualitative superiority.","section":"Section 4.6"},{"comment":"No model weights or code are released, and no statistical variation (e.g., multiple runs) is reported for the PSNR and runtime values; adding these would improve reproducibility.","section":"General"}],"recommendation":"reject","confidential_remarks":"The score discrepancy in Eq. (1) versus Tables 2 and 3 cannot be explained by a simple unit conversion, and the runtime source is contradictory between Sections 4.3 and 4.5. These are load-bearing issues for the paper's central claim of outperforming previous MAI champions, and correcting them would require additional measurements and possibly a re-evaluation of the architecture selection. I recommend rejection unless the authors can supply the official MAI scoring data and a full, consistent measurement protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, the paper reports composite scores that beat previous MAI champions, but the Score column in Tables 2 and 3 does not follow Equation (1). I checked: plugging the paper's own PSNR and runtime values into Score = 22·(PSNR−27)/runtime gives numbers 5–6 times larger than reported, and the ratios between models change. The claimed 66.7% and 88.3% improvements are not reproducible from the paper's data. Second, Section 4.3 says the MediaTek device was unavailable and evaluation was on a Snapdragon 870 GPU, yet Section 4.5 reports MediaTek NPU runtimes with no measurement details. That is a direct contradiction.\n\nOn the positive side: the architecture is a sensible combination of known building blocks — RepVGG-style reparameterization, FGNAS search, and a 1×1 conv before depth-to-space to cut upsampling cost. The ablation in Table 2 is at least internally consistent about PSNR and runtime trends, even if the Score column is wrong. The paper honestly cites prior work including the authors' own EVSRNet, and the PSNR numbers (27.79 dB) are plausible.\n\nThe soft spots are real. The score inconsistency is load-bearing because the paper's central claim is exactly that this model scores higher than prior champions. Without a corrected score column, that claim is unsupported. The hardware contradiction undermines the runtime input. The novelty is limited — an incremental engineering adaptation, not a new framework. The NAS-driven free parameters are optimized on the same score, so the headline score is a fitting result, not a prediction.\n\nWho this is for: readers tracking mobile AI challenge entries might browse it for the architectural recipe, but as a research paper it does not stand on its own evidence. I would not cite it given the current numbers. If the authors fix the score formula and clarify the runtime measurement, it could become a useful workshop report.\n\nRecommendation: I would not send this to peer review in its current form. The internal inconsistency is severe enough that a referee would reject or demand major revision. If the authors resubmit with corrected tables and a transparent measurement section, it could be reviewed as a modest engineering contribution. For now, desk reject is appropriate.","headline":"The architecture is a sensible engineering mix, but the score column does not match the paper's own formula, so the headline comparison is not reproducible.","tokens_in":8189,"tokens_out":3598,"would_cite":false,"duration_ms":30224,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"RepNet-VSR claims a higher MAI2025 composite score than previous champions with real-time mobile NPU runtime.","keywords":["video super-resolution","reparameterization","neural architecture search","mobile NPU","real-time inference","REDS dataset","model compression"],"falsifier":"Run the exported TensorFlow Lite model on a MediaTek Dimensity device under the official AI Benchmark protocol in FP16 with the MediaTek neuron delegate, measuring the time for 10 frames; if 22·(PSNR−27)/runtime falls below EVSRNet's 0.0174, the paper's central superiority claim fails.","tokens_in":7154,"feed_emoji":"📱","tokens_out":8018,"duration_ms":62481,"temperature":0.7,"pith_summary":"This paper is trying to establish that real-time 4× video super-resolution on mobile NPUs can be accurate enough to beat the previous champion algorithms in the MAI challenge series. The authors propose RepNet-VSR, a convolutional architecture that uses reparameterizable bottleneck blocks and a 1×1 convolution in place of the usual 3×3 channel-reduction layer before depth-to-space upsampling. They report 27.79 dB PSNR on the REDS validation set for 180p→720p upscaling, with a runtime of 103 ms per 10 frames on a MediaTek Dimensity NPU. Plugging those numbers into the MAI2025 score formula gives 0.029, higher than the EVSRNet and RCBSR scores of 0.0174 and 0.0154, which is the paper's central evidence that accuracy and efficiency can be co-optimized on edge hardware.","feed_headline":"Phone NPU video super-resolution hits 27.79 dB, tops MAI champs","feed_subtitle":"RepNet-VSR's reparameterized blocks and 1×1 upsampling bottleneck keep the composite score ahead of EVSRNet and RCBSR.","key_machinery":"RepConv is the load-bearing module: during training it is a bottleneck of a 1×1 expansion, a 3×3 convolution, a 1×1 reduction, plus a 1×1 residual shortcut; at inference these branches are merged into one 3×3 convolution so the deployed graph is as cheap as a plain convolutional block. The second component is the 1×1 convolution placed right before the depth-to-space (pixel-shuffle) layer, which cuts the channel count from 48 to 3 and avoids the expensive 3×3 upsampling bottleneck. A neural architecture search over channel count (0–32) and number of RepConv blocks (0–8), regularized by FLOPs, selects the configuration used for the final model, and a second phase fine-tunes that model with an L2 loss.","core_discovery":"The central claim is that the architecture, RepNet-VSR, obtains a better accuracy-efficiency trade-off on the MAI2025 scoring metric than the previous challenge winners. The key quantitative claim is a PSNR of 27.79 dB on REDS validation for 4× super-resolution (180p→720p) and a runtime of 103 ms per 10 frames on a MediaTek Dimensity NPU, giving a composite score of 0.029 versus 0.0174 for EVSRNet and 0.0154 for RCBSR. The authors attribute the gain to replacing the 3×3 convolution that compresses 48 channels to 3 before depth-to-space with a 1×1 convolution (which they say speeds up upsampling by a factor of 4 and improves quality), and to reparameterized RepConv blocks that train with a multi-branch structure and fold into a single 3×3 convolution at inference.","pith_inferences":["The paper's Section 4.3 says the required MediaTek hardware was unavailable and runtime was measured on a Qualcomm Snapdragon 870 GPU, while Section 4.5 states MediaTek NPU results without detailing the measurement; if the 103 ms figure was obtained differently, the composite score advantage could shrink or disappear.","The CPU runtime is 2.44× slower than RCBSR (273 ms vs 112 ms), so the reported advantage is specific to NPU execution; on systems where parts of the graph run on the CPU, RCBSR may be the better choice.","A natural test of generality would be to insert the same 1×1 upsampling bottleneck into a transformer-based VSR model and check whether the 4× upsampling speedup and quality retention replicate outside convolutional backbones."],"forward_implications":["If the reported runtime holds under the official protocol, 4× video super-resolution at 720p output becomes practical on current phone NPUs without cloud offloading.","The 1×1-before-depth-to-space substitution is a drop-in change for other super-resolution backbones and should cut upsampling cost whenever the pre-upsampling channel count is large.","The NAS search over channels and RepConv blocks yields a configurable trade-off curve, so the same architecture can be tuned for faster or higher-fidelity deployment.","On the MAI2025 score formula, the reported configuration beats both prior champions, making it a reference point for future mobile SR challenge entries."],"supporting_citations":[{"why":"Supplies the EVSRNet baseline whose composite score the paper claims to beat.","marker":"[32]"},{"why":"Supplies the RCBSR baseline with its reparameterized ECB block, the main comparison for both quality and runtime.","marker":"[34]"},{"why":"Provides the neural architecture search method used to select channel counts and RepConv block numbers.","marker":"[35]"},{"why":"Supplies the structural reparameterization technique that RepConv builds on.","marker":"[28]"},{"why":"Provides the REDS dataset used for training and validation.","marker":"[36]"},{"why":"Defines the AI Benchmark protocol referenced for the smartphone inference runtime measurement.","marker":"[1]"}],"fun_headline_variants":["Phone NPU runs 4x video super-res at 27.79 dB","RepNet-VSR: 4x super-res on mobile, beats MAI champs","Real-time 4x VSR with 1x1 upsampling: 103 ms per 10 frames","Reparameterized VSR tops MAI score, runs on phone NPU","27.79 dB super-res on REDS, 103 ms on NPU"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The 103 ms per 10 frames runtime on a MediaTek Dimensity NPU was produced by the official MAI2025 benchmarking procedure, even though the paper reports that the specified hardware was unavailable during the evaluation and does not describe how the final NPU measurement was made.","fun_headline_variants_meta":{"raw":{"variants":["Phone NPU runs 4x video super-res at 27.79 dB","RepNet-VSR: 4x super-res on mobile, beats MAI champs","Real-time 4x VSR with 1x1 upsampling: 103 ms per 10 frames","Reparameterized VSR tops MAI score, runs on phone NPU","27.79 dB super-res on REDS, 103 ms on NPU"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001095,"raw_usage":{"total_tokens":4560,"prompt_tokens":925,"completion_tokens":3635,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":3520}},"tokens_in":541,"tokens_out":3635,"duration_ms":24385,"temperature":1.0,"reasoning_tokens":3520,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:21:02.357533+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exported TensorFlow Lite model on a MediaTek Dimensity device under the official AI Benchmark protocol in FP16 with the MediaTek neuron delegate, measuring the time for 10 frames; if 22·(PSNR−27)/runtime falls below EVSRNet's 0.0174, the paper's central superiority claim fails.","supporting_citations":[{"cited_title":"Evsrnet: Efficient video super- resolution with neural architecture search","cited_arxiv_id":null,"evidence_quote":"Supplies the EVSRNet baseline whose composite score the paper claims to beat."},{"cited_title":"Power efficient video super-resolution on mo- bile npus with deep learning, mobile ai & aim 2022 challenge: Report","cited_arxiv_id":null,"evidence_quote":"Supplies the RCBSR baseline with its reparameterized ECB block, the main comparison for both quality and runtime."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the neural architecture search method used to select channel counts and RepConv block numbers."},{"cited_title":"Repvgg: Making vgg-style convnets great again","cited_arxiv_id":null,"evidence_quote":"Supplies the structural reparameterization technique that RepConv builds on."},{"cited_title":"Ntire 2019 challenge on video deblur- ring and superresolution: Dataset and study","cited_arxiv_id":null,"evidence_quote":"Provides the REDS dataset used for training and validation."},{"cited_title":"Ai benchmark: All about deep learning on smartphones in 2019","cited_arxiv_id":null,"evidence_quote":"Defines the AI Benchmark protocol referenced for the smartphone inference runtime measurement."}],"review_version":1}