{"id":"7e8957d3-c1af-4b3e-a931-0696e8ea9ecc","arxiv_id":"2412.01456","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A lightweight transformer that uses phase-only features as attention queries and keys achieves state-of-the-art underwater image restoration on several benchmarks.","lead":"This paper proposes Phaseformer, a 1.77 million parameter transformer that uses Fourier phase information in its attention mechanism to restore degraded underwater images. The authors report top scores on UIEB, UFO-120, and several real-world benchmarks, plus low-light enhancement results.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The SOTA claim rests on Table 1, which lists an impossible SSIM of 15.78 for RGHS on UIEB (SSIM is bounded by 1); until this table is corrected and independently reproduced, 'outperforms existing SOTA' is not verifiable.","rationale":"I read the paper as claiming two things: (i) phase-only features are less degraded and therefore useful for attention, and (ii) this yields SOTA numbers. The phase-invariance premise is supported only by a t-SNE figure and citations, which is a genuine weakness and the reader rightly flags it. However, the more load-bearing issue for the central claim is the reliability of the quantitative evidence. Table 1's RGHS SSIM of 15.78 is impossible and indicates that the comparison pipeline itself is compromised. If the table has errors, the claimed 'outperforms SOTA' collapses regardless of whether the phase hypothesis is true. The lack of released code and error bars means the reader cannot distinguish a real gain from evaluation noise. I therefore center the condition on reproducible numbers. A single concrete check, reproducing Table 1 with the authors' code and standard metrics on the official splits, would settle whether the concern lands. Since the reader already assigned CONDITIONAL largely for these reasons, my read does not move the verdict; it sharpens the condition.","tokens_in":13515,"tokens_out":5919,"duration_ms":53572,"concrete_test":"Request the authors' released code and trained weights, then rerun the UIEB evaluation on the official 90-image test split, computing PSNR and SSIM with standard implementations (e.g., scikit-image). Verify that Ours reproduces 25.98/0.928 and that every baseline SSIM in Table 1 is <= 1, especially RGHS where SSIM is entered as 15.78. Independently recompute all baseline numbers from released implementations using the identical split and augmentation protocol; if the reported RGHS value does not become a valid SSIM, or if Ours cannot be reproduced, Table 1 cannot support the SOTA claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim is that Phaseformer outperforms existing SOTA methods. The only quantitative support for that claim is Tables 1 and 2, and Table 1 contains an impossible value: RGHS on UIEB is reported with SSIM = 15.78, although SSIM is bounded above by 1. This is either a metric-computation bug or a transcription error in the very table used to establish the SOTA margin. The same table reports Ours at 25.98/0.928, but no code, trained weights, error bars, or significance tests are provided, so there is no way to tell whether the margin over Semi-UIR and Spectroformer is real or an artifact of the evaluation protocol. Section 4.2's split counts (3000/4800 training pairs for UFO-120/UIEB) are not reconciled with Section 4.1's stated 1500/800 training pairs; if the test split or augmentation differs from prior work, the comparisons are not apples-to-apples. Until these numbers are reconciled, the outperformance claim is unverified.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Phaseformer, a lightweight transformer for underwater image restoration with 1.77M parameters. The core idea is to use phase-only features, obtained by setting the FFT amplitude to one in a phase extraction module, as queries and keys in multi-head self-attention, and to propagate encoder features through an optimized phase attention block in the skip connections. The training loss combines Charbonnier, gradient, MS-SSIM, and perceptual terms with learnable per-loss weights plus fixed high/low-resolution weights. The method is evaluated on UIEB, UFO-120, U45, UCCS, and SQUID, and is also applied to low-light enhancement and downstream detection/depth tasks. The central claim is that Phaseformer outperforms existing state-of-the-art methods on these benchmarks.","tokens_in":1459,"tokens_out":3098,"duration_ms":50840,"significance":"If the empirical claims hold, the paper makes a useful contribution: it introduces a phase-based attention mechanism that is coherent, lightweight, and supported by ablation studies showing consistent gains from each proposed component. The authors provide an explicit parameter/FLOPs comparison, a code availability statement, and evaluate across several standard underwater datasets, which strengthens the practical relevance. The main significance lies in the potential of phase-only attention for restoration tasks, but this significance is currently limited by evaluation inconsistencies and the lack of a quantitative validation of the core phase-degradation premise.","major_comments":[{"comment":"The UIEB block of Table 1 reports SSIM = 15.78 for RGHS, which is impossible because SSIM is bounded above by 1. This is not a minor typo in isolation: Table 1 is the primary quantitative basis for the claim that Phaseformer outperforms existing SOTA methods. The authors must correct this value, recheck every other entry in Tables 1 and 2 with a standard evaluation script, and report the exact version of the metrics used, before the outperformance claim can be considered verified.","section":"Table 1, Section 4.4"},{"comment":"The dataset split counts are inconsistent across the two subsections. Section 4.1 states that UIEB uses 800 training pairs and 90 test pairs and that UFO-120 uses 1500 training pairs and 120 test pairs, while Section 4.2 states that augmentation produced 3000/4800 training image pairs and 120/90 testing image pairs for UFO-120/UIEB. The authors need to clarify whether these numbers refer to augmented copies, how many unique training images are used per dataset, and whether the test sets are identical to those used by the compared methods; otherwise the comparisons in Table 1 may not be on equal footing.","section":"Sections 4.1 and 4.2"},{"comment":"No error bars, standard deviations, or significance tests are reported for any metric. The claimed margins are small in several cases, e.g., SSIM 0.928 versus 0.917 for U-shape on UIEB, and the impossible RGHS SSIM value suggests the evaluation pipeline is not fully controlled. The authors should report multiple runs or at least provide released code and trained weights so that the margins can be independently reproduced.","section":"Tables 1 and 2"},{"comment":"The load-bearing premise that phase is less affected by underwater degradation than amplitude is supported only by the qualitative t-SNE plot in Figure 1 and by citations, including the authors' own prior work [34]. The paper does not quantify the degree to which phase is preserved under the actual degradation model used in training or evaluation, nor does it show that the learned attention maps indeed rely on phase information. Please add a quantitative experiment, such as phase/amplitude reconstruction error on degraded-clean pairs, or an attention-map analysis, to substantiate the motivation for the PEM module and to show that discarding amplitude information is not the sole source of the improvement.","section":"Sections 1 and 3.1"},{"comment":"The adaptive kernel-size formula is ambiguous as written: k = | log2(C')/gamma + b/gamma |_odd. In particular, it is unclear whether the denominator gamma applies to the whole numerator or only to log2(C'), and the placement of the nearest-odd operation relative to the division should be stated unambiguously. This makes the OPAB module difficult to reimplement exactly as described.","section":"Section 3.2, Eq. (5)"}],"minor_comments":[{"comment":"The method name 'Spectrofomer' is misspelled; it should be 'Spectroformer'.","section":"Table 1"},{"comment":"The text says 'Our method's performance aligns competitively with existing approaches,' which is much weaker than the abstract's claim that the approach 'outperforms existing SOTA methods.' Please make the wording consistent with the reported numbers.","section":"Section 4.4"},{"comment":"References [70] and [71] appear to refer to the same paper, 'Deep retinex decomposition for low-light enhancement.' One should be removed or the citation should be unified.","section":"References"},{"comment":"The code availability statement in the abstract says 'Code is available at Phaseformer,' but no URL or repository identifier is provided. Please include a working link or at least a complete reference to the repository.","section":"Abstract and Section 1"},{"comment":"The table header uses 'OPA' and 'OA' with checkmarks that are difficult to map to the five rows; please rename the columns or add a legend so the reader can identify which configuration is ablated in each row.","section":"Table 4"}],"recommendation":"major_revision","confidential_remarks":"The architectural idea is promising and the ablations are internally consistent, but the evaluation section needs serious work before publication: the impossible SSIM value in Table 1, the dataset split inconsistencies, and the lack of error bars all bear directly on the headline SOTA claim. I see this as a major revision rather than a rejection, because the problems appear fixable within the manuscript's scope and the phase-attention mechanism has enough support from the ablation study to warrant a revised version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper's central idea—use Fourier phase features as query and key in a transformer for underwater restoration—is a legitimate extension of the authors' earlier phase-transfer work, and the architecture (phase-based self-attention plus phase attention skip block, adaptive loss weights) is coherent. The ablations in Table 4 show each component adds a consistent bump, which is evidence the mechanism is doing something. The parameter count and FLOPs are genuinely low. So there is a real contribution here, in a modest subfield.\n\nBut the empirical support for the headline claim is not in good shape. Table 1, the table that establishes the SOTA margin, lists RGHS at SSIM 15.78 on UIEB. SSIM is bounded above by 1. That's either a metric bug or a transcription error, and it makes me want to re-check every number in that table before trusting 'outperforms SOTA.' The dataset split counts are also inconsistent: Section 4.1 says UIEB has 800 training pairs and UFO-120 has 1500; Section 4.2 says augmentation produced 4800 and 3000. Those may be consistent if augmentation multiplies the data, but the text doesn't say that, and the ambiguity matters because the comparison against prior work is only valid if the splits and augmentation match. There are no error bars or significance tests anywhere, and no code, so the margins over Semi-UIR and Spectroformer (about 1 PSNR) could easily be noise or evaluation-protocol artifacts.\n\nThe phase-robustness premise is plausible—Oppenheim and Lim is a strong citation—but the paper relies on a qualitative t-SNE plot and its own prior work to justify it. It doesn't independently quantify how much phase survives underwater degradation. That's a soft spot, though not a fatal one.\n\nThe overlap with the authors' prior work [34] and Spectroformer [35] should be stated more explicitly; right now the 'first attempt' claim is weakened by that context.\n\nWho is this for? Researchers working on underwater restoration or Fourier-phase deep learning methods. They'll get a useful idea and a decent baseline, but they'd want to see a corrected and reproducible version. I'd send it to peer review—the idea deserves referee time—but I would not accept the current manuscript. The impossible SSIM needs to be fixed, the dataset splits reconciled, and error bars or code provided. My own verdict is skeptical until those things happen.","headline":"Plausible phase-attention idea for underwater restoration, but the headline SOTA claim rests on a table with an impossible SSIM value, so the empirical case is not yet made.","tokens_in":14296,"tokens_out":2365,"would_cite":false,"duration_ms":20085,"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 1.77M-parameter transformer restores underwater images by attending to Fourier phase","keywords":["Phaseformer","underwater image restoration","phase-based attention","Fourier phase","lightweight transformer","low-light image enhancement","adaptive loss weighting"],"falsifier":"Measure FFT phase error and amplitude error between paired clean and degraded underwater images, for example on UIEB training pairs; if the phase error is comparable to the amplitude error, the paper's core premise fails. Alternatively, train Phaseformer with added phase noise on the inputs and observe whether restoration quality drops as fast as it does under amplitude noise, or compare against the same transformer with randomly shuffled phase keys.","tokens_in":13344,"feed_emoji":"🌊","tokens_out":3643,"duration_ms":34578,"temperature":0.7,"pith_summary":"The paper proposes Phaseformer, a lightweight transformer for restoring color and detail in degraded underwater images. Its central bet is that the Fourier phase of an image carries structural information that survives underwater degradation better than the amplitude does, so attention should be computed from phase-only queries and keys. The network pairs this phase-based self-attention with a phase-based skip block and trainable loss weights, and the authors report PSNR/SSIM of 25.98/0.928 on UIEB and 30.38/0.878 on UFO-120, above the compared methods, with leading UIQM on U45. They also show the same network improves low-light images and helps downstream detection and depth estimation.","feed_headline":"Phase-based attention restores underwater images at 1.77M parameters","feed_subtitle":"Fourier phase drives a lightweight transformer that beats heavier models on UIEB and UFO-120.","key_machinery":"The phase extraction module (PEM) is the load-bearing mechanism: it computes the FFT of input features, discards the amplitude by setting it to 1, keeps the phase, and returns phase-only features via inverse FFT. These phase-only features become the queries and keys of a phase-based multi-head self-attention block, producing a transposed attention map of size channels-by-channels. An optimized phase attention block applies the same phase-only idea in the skip connections, using global average pooling and an adaptively sized 1D convolution to weight prominent structural features. A fourth component, adaptive loss weighting, makes the weights of Charbonnier, gradient, MS-SSIM, and perceptual losses trainable rather than hand-tuned.","core_discovery":"The paper's claim is that a transformer can restore underwater images more effectively when its self-attention operates on phase-only features rather than on the raw features themselves. Concretely, a phase extraction module applies an FFT, sets the amplitude to 1, keeps the phase, and applies an inverse FFT; these phase-only maps serve as queries and keys, while values come from the input projection. The paper further claims that an optimized phase attention block in the skip connections propagates structural information from encoder to decoder without transferring degraded amplitude content, and that adaptively learned loss weights improve training over fixed weights. With 1.77 million parameters and 13.0 GFLOPs, the reported results on synthetic and real-world underwater benchmarks are consistently better than the compared state-of-the-art methods, and the same architecture improves low-light image enhancement.","pith_inferences":["A testable extension left implicit by the paper is whether the phase-robustness argument transfers to other amplitude-dominant degradations such as haze, rain, or blur; running the same architecture on standard dehazing and deraining benchmarks would isolate whether the gain is phase-specific or just the transformer backbone.","Because the phase-only query and key discard amplitude entirely, color-critical scenes might lose useful amplitude cues; a variant that re-injects a small fraction of amplitude into the attention would test whether the phase-only choice is optimal or merely sufficient.","The paper's t-SNE evidence for phase robustness is qualitative; a quantitative stress test would be to add controlled phase noise versus amplitude noise to degraded inputs and measure which noise type degrades restoration performance faster.","The multi-resolution output with adaptive loss weighting is a general training recipe that could be transplanted to other image restoration networks even if the phase-based attention is replaced."],"forward_implications":["If phase information is indeed the least-degraded part of an underwater image, then attention computed from phase-only queries and keys should restore structure without needing a large or deep network.","The reported parameter count and FLOPs suggest the architecture is light enough for onboard preprocessing on autonomous underwater vehicles, where computational budget is tight.","The same phase-based attention generalizes, per the paper, to low-light image enhancement, so the mechanism is not specific to water degradation.","Improved restoration transfers to downstream tasks: the authors demonstrate better fish detection with YOLO-v3 and better depth estimation with DPT on enhanced images.","Trainable loss weights remove a manual tuning step and, according to the ablation, outperform fixed weights on UIEB."],"supporting_citations":[{"why":"Supplies the classical result that the phase of an image carries structural information, which motivates using phase for queries and keys.","marker":"[51]"},{"why":"Supports the claim that phase is less affected by noise than magnitude, a premise for the phase extraction module.","marker":"[8]"},{"why":"Cited for phase being less affected by degradation than amplitude in signal processing, underpinning the t-SNE motivation.","marker":"[21]"},{"why":"The authors' own earlier phase-transfer work on underwater images, used as evidence that degradation mostly attacks amplitude.","marker":"[34]"},{"why":"Provides the UIEB benchmark dataset used for training and testing the central comparisons.","marker":"[39]"},{"why":"Provides the UFO-120 synthetic dataset used for the other main quantitative comparison.","marker":"[27]"},{"why":"The U-shape transformer is a key baseline and the work whose attention-based skip connection the optimized phase attention block is designed to improve upon.","marker":"[55]"},{"why":"Spectroformer is a recent transformer baseline that the paper compares against on UIEB, U45, and complexity metrics.","marker":"[35]"}],"fun_headline_variants":["Phase-only attention clears murky water at 1.77M params","Fourier phase powers tiny transformer to beat SOTA underwater","Lightweight phase transformer excels at underwater and low-light restoration","Phase-based self-attention: key to underwater image restoration and beyond"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole design leans on the claim that underwater degradation attacks the Fourier amplitude of an image far more than its phase; if real-world scenes corrupt phase substantially, the attention mechanism loses its proposed advantage.","fun_headline_variants_meta":{"raw":{"variants":["Phase-only attention clears murky water at 1.77M params","Fourier phase powers tiny transformer to beat SOTA underwater","Lightweight phase transformer excels at underwater and low-light restoration","Phase-based self-attention: key to underwater image restoration and beyond"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000239,"raw_usage":{"total_tokens":1483,"prompt_tokens":880,"completion_tokens":603,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":496,"completion_tokens_details":{"reasoning_tokens":530}},"tokens_in":496,"tokens_out":603,"duration_ms":5375,"temperature":1.0,"reasoning_tokens":530,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:17:38.985547+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure FFT phase error and amplitude error between paired clean and degraded underwater images, for example on UIEB training pairs; if the phase error is comparable to the amplitude error, the paper's core premise fails. Alternatively, train Phaseformer with added phase noise on the inputs and observe whether restoration quality drops as fast as it does under amplitude noise, or compare against the same transformer with randomly shuffled phase keys.","supporting_citations":[{"cited_title":"SIAM (2005) 3, 4","cited_arxiv_id":null,"evidence_quote":"Supports the claim that phase is less affected by noise than magnitude, a premise for the phase extraction module."},{"cited_title":"John Wiley & Sons (2008) 1, 3, 4","cited_arxiv_id":null,"evidence_quote":"Cited for phase being less affected by degradation than amplitude in signal processing, underpinning the t-SNE motivation."},{"cited_title":"In: 2023 International Joint Conference on Neural Networks (IJCNN)","cited_arxiv_id":null,"evidence_quote":"The authors' own earlier phase-transfer work on underwater images, used as evidence that degradation mostly attacks amplitude."},{"cited_title":"IEEE Transactions on Image Processing 29, 4376–4389 (2019) 2, 5","cited_arxiv_id":null,"evidence_quote":"Provides the UIEB benchmark dataset used for training and testing the central comparisons."},{"cited_title":"In: Computer Vision–ECCV 2022 Workshops: Tel Aviv, Israel, Oc- tober 23–27, 2022, Proceedings, Part II","cited_arxiv_id":null,"evidence_quote":"The U-shape transformer is a key baseline and the work whose attention-based skip connection the optimized phase attention block is designed to improve upon."},{"cited_title":"In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision","cited_arxiv_id":null,"evidence_quote":"Spectroformer is a recent transformer baseline that the paper compares against on UIEB, U45, and complexity metrics."}],"review_version":1}