{"id":"e1a35a63-beb8-465c-9753-277cdb08de25","arxiv_id":"2506.15929","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A RAW-to-sRGB demoireing model built from linear-attention blocks and a truncated flow-matching refinement step reports state-of-the-art PSNR and SSIM on two benchmarks, with internal reporting inconsistencies.","lead":"This paper combines a supervised deep network and a flow-matching generative prior to remove moiré patterns from RAW camera images and videos. The authors report higher PSNR and SSIM than previous demoireing methods, but the paper's own tables and text contain conflicting numbers and the 'test-time training' component is never actually trained at test time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The TFMP refinement, which provides the claimed SOTA gain, is unsupported: the flow-matching prior is undisclosed and the text/table give contradictory LPIPS (0.0795 vs 0.0973), so the +0.09 dB PSNR improvement cannot be verified.","rationale":"I agree with the reader's REJECT, emphasizing the TFMP refinement as the load-bearing element. The +0.09 dB gain is the entire justification for 'state-of-the-art' beyond the base MoiréXNet, and the paper's own evidence for it is contradictory: the text's LPIPS 0.0795 versus Table I's 0.0973, plus ablation increments that don't match Table III. The undisclosed prior makes the experiment irreproducible, and the omission of any data-fidelity term means the claimed MAP derivation does not justify the update rule. These are internal inconsistencies, not differences from community consensus. The base model may be competitive, but the submitted evidence does not support the new-benchmark claim, so the verdict should remain REJECT unless the authors provide checkpoints and corrected, consistent tables.","tokens_in":16879,"tokens_out":7937,"duration_ms":79778,"concrete_test":"Obtain the exact TFMP velocity-field checkpoint and the MoiréXNet checkpoint, then rerun the RawVDemoire test split with the same schedule (15 iterations, t=0.95) and recompute PSNR, SSIM, LPIPS. If the reproduced TFMP LPIPS is ≥0.0847 (matching Table I's 0.0973 degradation) or the checkpoint is unavailable, the refinement claim fails; if LPIPS drops to 0.0795 as the text says and PSNR gains +0.09 dB, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on the TFMP refinement in Section III-B, but that component is both underspecified and internally contradictory. The paper asserts that a pretrained flow-matching network 'maps degraded images to clean ones' and that integrating v(xt,t) from t=0.95 refines MoiréXNet outputs, yet it never states the prior's training data, architecture, or checkpoint, and the only cited prior (PnP-Flow [43]) is described in Fig. 1 as producing artifacts on moiré sRGB inputs. The update xt-1 = xt + Δt·v(xt,t) contains no data-fidelity or likelihood term, so it is not a MAP descent step for the model y=M(x)+n described in Section III; it is unconditional generative sampling, which can drift away from the observed frame. The reported evidence is self-contradictory: Section IV-D says TFMP reaches LPIPS 0.0795 (0.0054 below DemMamba's 0.0901), while Table I reports TFMP LPIPS as 0.0973, worse than MoiréXNet's 0.0847. If the table is correct, TFMP worsens perceptual quality; if the text is correct, the table is wrong. The ablation text in Section IV-E also gives PSNR increments (+0.99 for INN, +0.09 for LFEF) that do not match Table III (+0.32, +0.76). The plus-0.09 dB PSNR increment from TFMP is therefore unverifiable, and the paper's own numbers undermine the claim that TFMP 'further enhances' the base model.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MoiréXNet, a RAW-to-sRGB demoiréing model that adapts the VDRaw framework by inserting invertible neural network (INN) blocks, a Learnable Frequency Enhanced Filter (LFEF), and linear-attention TTT modules, and combines this supervised model with a Truncated Flow Matching Prior (TFMP) refinement stage. The authors claim a hybrid MAP-based formulation, state-of-the-art PSNR/SSIM on the RawVDemoiré dataset (30.214 dB for TFMP vs. 30.004 dB for DemMamba), competitive results on TMM22, and an inference time of 0.070 s. The two main novelties are the architectural modifications to VDRaw and the TFMP refinement.","tokens_in":17269,"tokens_out":9022,"duration_ms":85272,"significance":"If the base-model numbers are reliable, the VDRaw-with-TTT adaptation is a credible empirical contribution: it reports 30.127 dB video PSNR and 0.070 s inference, surpassing DemMamba while being much faster, and Table II shows competitive TMM22 performance. The paper gives a reasonably detailed architecture description and reports efficiency transparently. However, the TFMP component is the paper's distinctive claimed contribution and it is currently neither derived, specified, nor consistently reported: the update rule is not MAP, the pretrained velocity field is undisclosed, and the LPIPS values in the text and Table I contradict each other. The significance of the paper as a hybrid MAP framework therefore cannot be assessed in its present form.","major_comments":[{"comment":"The MAP framing is not established. Eq. (1) is written for the linear model y = Hx + n, while the paper's own problem statement in Section III is y = M(x) + n with nonlinear M. The TFMP update xt-1 = xt + Δt·v(xt,t) contains no data-fidelity or likelihood term with respect to y, so the claim that it optimizes the MAP objective in Eq. (1) is unsupported; it is an unconditional generative step that can drift away from the observed frame. Moreover, the pretrained flow-matching velocity field is never described: its training data, architecture, loss, and checkpoint are absent, and the only related prior cited (PnP-Flow [43]) is said in Fig. 1 to produce artifacts on moiré sRGB inputs. Without this information the +0.087 dB gain in Table I cannot be verified or reproduced.","section":"Section II-B / Section III-B"},{"comment":"The LPIPS reporting for TFMP is internally contradictory. Section IV-D states that the refined model 'achieves the lowest LPIPS score of 0.0795, which is 0.0054 lower than DeMMamba (0.0901)', but Table I lists TFMP LPIPS as 0.0973, which is worse than both MoiréXNet (0.0847) and DemMamba (0.0901). If the table is correct, the TFMP refinement degrades perceptual quality; if the text is correct, the table is wrong. This contradiction bears directly on the claim that TFMP 'further enhances' the base model, so the reported gain cannot be taken at face value.","section":"Table I / Section IV-D"},{"comment":"The ablation text does not match the ablation table. The text says adding INN gives a PSNR increase of +0.99, but Table III shows 29.04 to 29.36, i.e., +0.32; it says LFEF contributes +0.09, but the table shows 29.36 to 30.12, i.e., +0.76. Only the TFMP increment agrees (+0.09, from 30.12 to 30.21). Because this ablation is the evidence that the proposed architectural components are responsible for the reported gains, the mismatch must be resolved before the architecture contribution can be evaluated.","section":"Section IV-E / Table III"},{"comment":"The truncation parameters appear to be selected on the test set. Section IV-E states that 'the PSNR peaks around iteration 15' in Figure 5 and then sets t = 0.95 'to avoid overshooting the peak'. Since the refinement curve itself is used to pick the hyperparameters, the reported TFMP improvement is a selected result rather than an independent evaluation. In addition, Figure 5's vertical axis is labeled 'PSNR' but the plotted values are between 0 and 6; if these are PSNR values, they are implausible for the reported 30 dB range, and the axis needs correction or relabeling.","section":"Section IV-E / Figure 5"}],"minor_comments":[{"comment":"The Introduction uses 'TFPM' instead of 'TFMP', and Section IV-D calls the method 'PnP flow matching'; the terminology should be made consistent throughout.","section":"Section I, contribution 3"},{"comment":"Section IV-A says training begins with 'L1 VGG loss' and later fine-tunes with 'wavelet loss', but Eq. (2) only defines L1 and VGG losses, and the wavelet loss is never defined.","section":"Section IV-A / Eq. (2)"},{"comment":"SSIM is a unitless index, but the text reports 'SSIM is +0.0231 dB higher' and '+0.0141 dB higher'; the 'dB' unit should be removed.","section":"Section IV-D"},{"comment":"The sentence 'we set xt = x~, with t starting from a higher value (e.g., t = 0.95)' and the mention of 'five samples drawn at each step' are unclear; the paper does not explain what the five samples are or how they are aggregated into the update.","section":"Section III-B"},{"comment":"There are several typographical errors, including 'MoNoiréXNet' in Section IV-D, 'VDraw' in Section IV-E, and 'chanllenge' and 'demoi´ering' in Section II-B.","section":"Throughout"},{"comment":"The term 'Test-Time Training' is used for the TTT blocks, but no self-supervised test-time adaptation objective is described; the paper should clarify whether these are used simply as linear-attention feature extractors.","section":"Section III-A"}],"recommendation":"reject","confidential_remarks":"For the editor: the supervised backbone may be salvageable as a separate contribution, but the manuscript's stated core novelty, the TFMP/MAP refinement, is not reproducible as written and its reported numbers are internally inconsistent. A resubmission that either supplies the full prior specification and fixes the contradictions, or refocuses the paper on the supervised architecture alone, would be worth considering. The current version also shows signs of rushed preparation, including placeholder template text in the acknowledgments and several typographical slips."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my read of arXiv:2506.15929. The paper combines a VDRaw-style RAW-to-sRGB demoireing backbone with three additions: invertible neural network feature extraction, a learnable frequency filter, and linear attention blocks from the TTT family, then adds a truncated flow-matching refinement. The base engine is a reasonable engineering extension: replacing VDRaw's preprocessing with INN+LFEF+TTT blocks is coherent, and the reported inference time (0.07s) with competitive PSNR on RawVDemoire is genuinely interesting if the numbers hold. The authors also correctly identify that PnP-Flow, designed for linear inverse problems, does not transfer directly to moiré removal.\n\nThe soft spots are serious and, as the stress test notes, partly contradictory. The TFMP refinement is underspecified: the pretrained velocity field's training data, architecture, and checkpoint are never stated, and the update xt-1 = xt + Δt·v(xt,t) contains no data-fidelity term. So the MAP framing in Section III is asserted, not derived. More damagingly, the numbers do not agree. Section IV-D's text says TFMP reaches LPIPS 0.0795, but Table I reports 0.0973—worse than the base model. The ablation increments in the text (+0.99 for INN, +0.09 for LFEF) do not match Table III (+0.32, +0.76). And t=0.95 with the 15-iteration cap is chosen after inspecting the PSNR-versus-iteration curve on the evaluation set—selection on the test data. With no code or checkpoints, the +0.09 dB refinement gain cannot be verified.\n\nWhat is actually solid: the base model's architecture is a plausible adaptation, and the computational-efficiency claim is important for a practical task. But the core novelty—the TFMP refinement—is both undisclosed and unsupported by the paper's own tables. The \"TTT\" name also oversells: the blocks are used as parametric feature extractors, not as a test-time training procedure.\n\nMy recommendation: send it to peer review, but tell the reviewers to check the tables against the text, demand TFMP details, and require the authors to release the prior and code. If those issues are resolved, the base model contribution may stand; as written, the SOTA claim rests on a component that cannot be reproduced.","headline":"A promising base model for RAW demoireing, undermined by an unverifiable refinement stage and internal numerical contradictions.","tokens_in":17832,"tokens_out":2616,"would_cite":false,"duration_ms":27532,"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":"MoiréXNet claims a new state of the art for RAW image and video demoiréing, reaching 30.214 dB PSNR on RawVDemoire video by combining linear-attention test-time training with a truncated flow-matching refinement.","keywords":["moire pattern removal","RAW-to-sRGB demoiréing","test-time training","linear attention","flow matching prior","video restoration","frequency domain filtering","invertible neural networks"],"falsifier":"Re-run MoiréXNet on RawVDemoire with the reported t=0.95 and 15 refinement iterations, and record PSNR and LPIPS before and after flow matching; the table lists TFMP LPIPS as 0.0973 while the text claims 0.0795, so resolving this discrepancy and testing several independent velocity fields would show whether the refinement actually improves perceptual quality or merely shifts pixel metrics.","tokens_in":16612,"feed_emoji":"🖼️","tokens_out":7607,"duration_ms":77972,"temperature":0.7,"pith_summary":"The paper proposes MoiréXNet, a hybrid demoiréing system that removes moiré patterns directly from RAW camera data and outputs clean sRGB images. Its central claim is that combining a fast supervised network built on linear-attention test-time training with a truncated flow matching prior yields better restoration than either approach alone, because moiré degradation is nonlinear and does not fit the linear forward models most generative restoration assumes. On the RawVDemoire benchmark the paper reports 29.590 dB PSNR for images and 30.127 dB for video, rising to 30.214 dB after the flow matching refinement, with inference at 0.070 seconds per frame. A sympathetic reader would care because the method addresses both quality and speed for a practical capture artifact, and because the same hybrid could extend to other nonlinear degradations.","feed_headline":"One demoiréing model hits 30.214 dB at 0.07 seconds per frame","feed_subtitle":"A hybrid of linear-attention test-time training and a flow-matching prior beats prior RAW demoiréing methods on both speed and quality.","key_machinery":"The load-bearing machinery is the MAP-style marriage of two components. The first is a supervised RAW-to-sRGB network whose deep feature extractor stacks linear-attention TTT blocks, each maintaining a compact parametric hidden state updated as $s_t = f(s_{t-1}, x_t; W)$, so memory stays fixed instead of growing with sequence length; before these blocks, an invertible neural network preserves information losslessly and a learnable frequency enhanced filter reweights low- and high-frequency components. The second component is the Truncated Flow Matching Prior, a generative velocity field $\\partial x_t/\\partial t = v(x_t,t)$ learned to map degraded images to clean ones, applied only near the clean end of the trajectory by setting $x_t$ to the supervised output and integrating $x_{t-1} = x_t + \\Delta t\\, v(x_t,t)$ from $t=0.95$ for a few iterations. Multi-scale features, pyramid cascading deformable alignment for video, and multiscale reconstruction losses tie the stages together.","core_discovery":"MoiréXNet claims to establish a new benchmark for RAW image and video demoiréing. The supervised stage maps three neighboring RAW frames directly to the sRGB central frame using multi-scale features, invertible lossless transforms, a learnable frequency filter, and TTT linear attention blocks; it alone beats prior RAW and sRGB methods on RawVDemoire in PSNR, SSIM, LPIPS, and inference speed. The truncated flow matching prior then integrates a pretrained velocity field for about 15 steps from t = 0.95 rather than t = 1, nudging the supervised output toward the clean image distribution and adding roughly 0.09 dB PSNR. The paper also reports competitive results on the TMM22 RAW image dataset, where MoiréXNet matches or slightly trails the strongest Mamba-based rival on PSNR while improving LPIPS.","pith_inferences":["The refinement step is only as good as the externally pretrained velocity field, which the paper never specifies; a testable extension would train the flow matching prior on the same RAW demoiréing task and compare gains.","If the reported speed holds, the same TTT linear-attention backbone could be transferred to other nonlinear, spatially varying degradations such as reflection removal or JPEG artifact reduction, where plug-and-play generative priors are also known to struggle.","The LPIPS discrepancy between Table I and the text suggests the perceptual effect of flow matching refinement is unstable; measuring LPIPS before and after refinement on a held-out split would clarify whether the generative step genuinely helps or trades pixel fidelity for artifacts.","Because TFMP starts at t=0.95 and runs only about 15 iterations, it could be viewed as a lightweight post-processing module rather than a full generative restoration, allowing it to be swapped in or out of a deployment pipeline without retraining the supervised model."],"forward_implications":["On RawVDemoire, MoiréXNet reports image demoiréing at 29.590 dB PSNR, video demoiréing at 30.127 dB PSNR, and video SSIM of 0.9258, all above the listed baselines and at 0.070 seconds per frame.","Adding the truncated flow matching prior raises the video numbers to 30.214 dB PSNR and 0.9281 SSIM according to the paper, while the reported LPIPS moves in opposite directions depending on whether one reads the table or the text.","Because TTT linear attention uses a fixed-size hidden state rather than an explicit key-value cache, the architecture's memory cost is $O(1)$ per sequence, which is what makes the fast multi-scale video inference possible.","The ablation attributes the bulk of the quality gain to the invertible and frequency-filter modules, with the flow matching prior contributing a smaller increment."],"supporting_citations":[{"why":"VDRaw supplies the base video demoiréing architecture and the RawVDemoire evaluation setup that MoiréXNet inherits and modifies.","marker":"[63]"},{"why":"DemMamba is the strongest prior video demoiréing baseline, with 30.004 dB PSNR, which MoiréXNet and TFMP claim to surpass.","marker":"[78]"},{"why":"RDNet and the TMM22 dataset provide the RAW image demoiréing benchmark and baseline for comparison.","marker":"[10]"},{"why":"RawVDemoire is cited as the RAW video demoiréing dataset with temporal alignment methods.","marker":"[15]"},{"why":"PnP-Flow defines the plug-and-play flow matching restoration approach whose learned velocity field TFMP adapts.","marker":"[43]"},{"why":"The TTT literature supplies the test-time training hidden-state mechanism used in the linear attention blocks.","marker":"[51]"},{"why":"The invertible neural network used for lossless shallow feature extraction is taken from this work.","marker":"[80]"},{"why":"RRID is the RAW+sRGB image demoiréing baseline that MoiréXNet outperforms in Table I.","marker":"[59]"},{"why":"VDmoire supplies a video demoiréing dataset and baseline compared in Table I.","marker":"[14]"}],"fun_headline_variants":["MoiréXNet: 30.214 dB PSNR at 0.07 s per frame","Hybrid TTT and flow prior demoiréing hits 30.214 dB","Faster and sharper demoiréing: 30.214 dB, 0.07 s per frame","MoiréXNet beats prior RAW demoiréing with TTT plus flow prior"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a pretrained flow-matching velocity field, whose training data, architecture, and checkpoint the paper never states, maps MoiréXNet's outputs on the test distribution toward clean images when integrated from t=0.95; if that field was trained on different degradations, the claimed 0.09 dB refinement gain could disappear or reverse.","fun_headline_variants_meta":{"raw":{"variants":["MoiréXNet: 30.214 dB PSNR at 0.07 s per frame","Hybrid TTT and flow prior demoiréing hits 30.214 dB","Faster and sharper demoiréing: 30.214 dB, 0.07 s per frame","MoiréXNet beats prior RAW demoiréing with TTT plus flow prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000445,"raw_usage":{"total_tokens":2261,"prompt_tokens":966,"completion_tokens":1295,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":582,"completion_tokens_details":{"reasoning_tokens":1193}},"tokens_in":582,"tokens_out":1295,"duration_ms":11566,"temperature":1.0,"reasoning_tokens":1193,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:45:32.390038+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run MoiréXNet on RawVDemoire with the reported t=0.95 and 15 refinement iterations, and record PSNR and LPIPS before and after flow matching; the table lists TFMP LPIPS as 0.0973 while the text claims 0.0795, so resolving this discrepancy and testing several independent velocity fields would show whether the refinement actually improves perceptual quality or merely shifts pixel metrics.","supporting_citations":[{"cited_title":"Recaptured raw screen image and video demoireing via channel and spatial modulations,","cited_arxiv_id":null,"evidence_quote":"VDRaw supplies the base video demoiréing architecture and the RawVDemoire evaluation setup that MoiréXNet inherits and modifies."},{"cited_title":"Cddfuse: Correlation-driven dual-branch feature decomposition for multi-modality image fusion,","cited_arxiv_id":null,"evidence_quote":"The invertible neural network used for lossless shallow feature extraction is taken from this work."},{"cited_title":"Image demoireing in raw and srgb domains,","cited_arxiv_id":null,"evidence_quote":"RRID is the RAW+sRGB image demoiréing baseline that MoiréXNet outperforms in Table I."}],"review_version":1}