{"id":"1d44ad16-dc65-4ccc-9fb1-17c619a0a908","arxiv_id":"2507.22501","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A degradation score trained on PSNR conditions a Swin UNet diffusion model for underwater enhancement, reporting top UIEB and LSUI scores, but the reference-free inference mechanism is unspecified.","lead":"DACA-Net is a deep learning system that estimates how badly an underwater photo is degraded, then uses that estimate to guide a diffusion model that restores color and detail. It matters because underwater cameras are used for marine monitoring and robotics, and the paper reports top scores on two standard underwater enhancement benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The degradation score that conditions the entire DACA-Net pipeline is defined in Eq. (1) as a function of both raw and reference images; no test-time substitute is described, so the reported enhancement cannot be reproduced on real inputs.","rationale":"Read in good faith, the paper's proposed contribution is a plausible-sounding conditional diffusion framework, but the mechanism is specified incompletely. The reader's strongest claim is that the method achieves SOTA due to degradation-aware conditioning. The weakest assumption is exactly the availability of D at inference. I agree with that identification. Section 3.2 is explicit: D is a function of both I_raw and I_ref, supervised by reference-based PSNR. Nothing in the paper describes a single-image degradation predictor at test time. This is not a consensus disagreement or a missing experiment; it is an internal inconsistency between the architecture as specified and the deployment scenario claimed in the abstract. The diffusion training objective gap (starting from x_T = raw image without a stated forward/noising loss) reinforces the concern but is secondary. I also note the paper provides no code, no error bars, and no random seeds; however, the decisive issue is logical, not empirical. For the central claim to hold, one must be able to compute D from I_raw alone; since Eq. (1) and Table 1 tie D to the reference, the burden is on the authors to supply the missing test-time mechanism. My recommendation is unchanged relative to the reader's verdict: reject as stated, because the main contribution cannot be instantiated. The paper contains some useful components (dual-stream estimator results, ablations), but those do not rescue the central claim.","tokens_in":14197,"tokens_out":3988,"duration_ms":44910,"concrete_test":"Obtain the released code/checkpoint (the paper provides none) or ask the authors for the exact test-time computation of D. Trace the forward pass on a single UIEB test image with the reference withheld. If D requires two inputs, the pipeline is not deployable; alternatively, rerun Table 3 with D fixed to a constant (e.g., 0) or predicted by a single-stream network trained on I_raw alone. If PSNR/SSIM changes materially from 28.60 dB / 0.9456, the reported advantage depends on a conditioning signal that real inputs cannot provide.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim — that degradation-aware conditioning drives the reported gains — is not instantiable at inference. Eq. (1) defines the conditioning signal as D = R(phi(I_raw), phi(I_ref)), a function of the degraded image and its reference. Table 1 evaluates this estimator on paired inputs, with ground-truth labels from PSNR via Eq. (2), and Section 3.1 says the score is regressed 'from paired raw and reference images.' No test-time procedure for computing D from I_raw alone is given. The abstract and Figure 1 do say the raw image is the input, but that is inconsistent with Eq. (1) unless a surrogate exists. The score enters the adaptive noise schedule (Eq. 5), AdaGN (Eq. 6), and PGFM red-channel gain (Eq. 7), so its absence removes the distinguishing mechanism; the method degenerates to an unspecified unconditional or constant-conditioned diffusion model. Table 3 and the ablations in Tables 4-5 all assume D is available, so the headline results are not reproducible as stated. A related gap compounds this: Section 3.3 starts the reverse process from the raw image embedded as x_T, yet no forward noising scheme or denoising objective for this nonstandard start state is specified. The missing inference-time D alone is sufficient to break the central claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DACA-Net proposes a degradation-aware conditional diffusion model for underwater image enhancement. A dual-stream estimator regresses a scalar degradation score D from the degraded image and its reference, supervised by normalized PSNR; D is injected into a Swin-UNet diffusion backbone through an adaptive noise schedule, AdaGN, and a physical-guided fusion module. Training uses a hybrid loss of perceptual, histogram, and contrastive terms. On UIEB and LSUI the paper reports improvements over physics-based and learning-based baselines. The central problem is that D is defined on paired inputs, so the method is not deployable at inference as described, and the diffusion training objective is not specified.","tokens_in":14503,"tokens_out":8292,"duration_ms":95038,"significance":"If the proposed conditioning mechanism worked, a continuous degradation score could provide a principled way to adapt diffusion-based restoration to varying underwater degradations. The paper does not ship code or trained models, and all claimed results come from a single run without error bars, so the empirical contribution is not independently verifiable. The degradation-score idea is plausible, but the manuscript does not establish a working instance of it.","major_comments":[{"comment":"The degradation score is defined as D = R(phi(I_raw), phi(I_ref)), and Section 3.1 states it is regressed 'from paired raw and reference images.' No test-time procedure is given for computing D from the raw image alone, even though the abstract and Figure 1 describe the raw image as the input. Since D modulates the noise schedule (Eq. 5), AdaGN (Eq. 6), and PGFM red-channel gain (Eq. 7), the method as written cannot be deployed on real underwater images; the headline results in Tables 3-5 therefore cannot be reproduced from the paper alone.","section":"Section 3.2, Eq. (1)"},{"comment":"Standard DDPM training assumes x_T ~ N(0,I) and a forward noising process on clean images. The paper instead states that 'the input is a raw underwater image, which is embedded as x_T and denoised,' but it never specifies the forward noising scheme, the noise-prediction objective, or any training loss for the reverse process from a non-Gaussian start state. Without this objective, the claimed conditional diffusion mechanism is undefined.","section":"Section 3.3, Eqs. (3)-(4)"},{"comment":"The total loss L_total consists only of perceptual, histogram, and contrastive terms; there is no diffusion denoising loss (e.g., an epsilon-prediction or variational bound term). It is therefore unclear what objective trains the denoising network epsilon_theta, and the paper does not explain how the 'progressive denoising' contributes beyond the three loss terms.","section":"Section 3.5, Eq. (12)"},{"comment":"The text states that DACA-Net obtains the highest UCIQE on UIEB, but the table lists DM_underwater with UCIQE 0.7230 versus DACA-Net's 0.7138, so the claim is contradicted by the paper's own numbers. More generally, all reported numbers are from a single run with no error bars, seeds, or code, which is insufficient support for the state-of-the-art claim.","section":"Section 4.3.3, Table 3"},{"comment":"The ablation narrative reports that removing all three loss components yields PSNR 15.08, but no row in Table 5 contains that value, and the row formatting makes the component configuration unreadable. The claimed baseline and the ablation conclusions are thus not verifiable.","section":"Section 4.4.1, Table 5"}],"minor_comments":[{"comment":"The parameter count column is labeled 'Param.(G)' and reports 1.02 for the proposed estimator; if the unit is gigaparameters this contradicts the 'lightweight' claim, and if it is megaparameters the header should be corrected.","section":"Table 1"},{"comment":"The symbol F is used both for the feature map and for the FFT operator, making the formula difficult to parse.","section":"Eq. (8)"},{"comment":"The direction of D should be stated explicitly: because D = 1 - normalized PSNR, higher D corresponds to stronger degradation; this is not obvious when it appears as a multiplicative factor in Eqs. (5) and (7).","section":"Section 3.2, Eq. (2)"},{"comment":"The caption says the method achieves 'best or second-best performance,' but the text claims it is best on all metrics; on UIEB UCIQE the table shows DM_underwater ahead of DACA-Net.","section":"Table 3 caption"},{"comment":"The inference-time measurement of the estimator is not accompanied by the hardware or resolution at which the time is measured; please specify the experimental conditions.","section":"Section 4.3.1"}],"recommendation":"reject","confidential_remarks":"The manuscript appears to be an ACM MM submission. The main technical gap—the reference-dependent degradation score—is not a presentation issue but a fundamental obstacle to using the method. If the authors can provide a test-time estimator and a proper diffusion training objective, a substantially revised manuscript could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the degradation score that conditions the entire pipeline is defined in Eq. (1) as a function of both the raw image and its reference. No test-time substitute is described anywhere. So as written, the method cannot produce a result from a single degraded image. That is a load-bearing gap, not a nitpick.\n\nWhat is genuinely new: the specific combination of a PSNR-regressed conditioning signal, an adaptive noise schedule (Eq. 5), AdaGN, and the physical-guided fusion block. I haven't seen that exact package before. The ablations in Tables 4 and 5 are thorough and show each component contributes; the degradation-estimator comparison in Table 1 is also useful. Credit where due: the authors did run real comparisons and report numbers.\n\nThe soft spots are serious. First, the missing inference-time D. The abstract and Figure 1 say the raw image is the input, but Section 3.1 says the score is regressed 'from paired raw and reference images,' and Eq. (1) makes the reference an explicit input. Table 1 evaluates the estimator on paired inputs with PSNR-derived labels. Without a surrogate for D computed from I_raw alone, the adaptive noise schedule, the AdaGN conditioning, and the red-channel gain (Eq. 7) all become undefined. The method degenerates to an unconditional or constant-conditioned diffusion model, and the reported gains cannot be reproduced. This is the same issue the stress-test note flags, and it holds up on reading.\n\nSecond, the diffusion formulation is not fully specified. Section 3.3 says the raw image is 'embedded as x_T' and then denoised, but the paper does not state the forward noising process or the training objective for that nonstandard start state. Standard DDPM starts from Gaussian noise; using the image as x_T is a different paradigm that needs its own justification.\n\nThird, the empirical claims come from a single run, with no error bars and no code. The margins over GUPDM (0.41 dB on UIEB, 1.15 dB on LSUI) are small enough that run-to-run variance could matter. The degradation score is calibrated to PSNR, which is also the evaluation metric, so the conditioning signal is partly circular.\n\nWho gets value: researchers working on conditional diffusion for restoration will find this a useful cautionary example of how a conditioning mechanism must be instantiable at inference. It is not a paper I would build on or cite in its current form.\n\nFor peer review: I would send it to reviewers, because the flaw is non-obvious and the empirical work is substantial. The right outcome is major revision with a clear test-time procedure for D, or a reformulation that drops the reference-based conditioning. As it stands, the central claim does not hold as written.","headline":"The degradation score that drives the whole pipeline needs the reference image at inference, so the method as written cannot be applied to a real underwater photo.","tokens_in":15005,"tokens_out":2860,"would_cite":false,"duration_ms":30192,"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":"DACA-Net claims that conditioning a diffusion denoiser on a continuous, PSNR-derived degradation score yields state-of-the-art underwater image enhancement: 28.60 dB PSNR on UIEB and 28.57 dB on LSUI, with ablations showing that the…","keywords":["underwater image enhancement","conditional diffusion model","degradation score","adaptive group normalization","physical-guided fusion","PSNR regression","Swin UNet","hybrid loss"],"falsifier":"Run the trained DACA-Net on a held-out underwater image with the reference withheld and trace whether the degradation estimator can produce D without the reference; if Eq. (1) is used as written, the conditioning signal is unavailable at inference and the claimed adaptive mechanism cannot operate. Alternatively, ablate D to a constant during denoising: if PSNR does not drop measurably, the degradation score is not carrying the improvement.","tokens_in":13986,"feed_emoji":"🌊","tokens_out":4416,"duration_ms":46374,"temperature":0.7,"pith_summary":"The paper tries to establish that underwater image enhancement can be made adaptive by feeding the diffusion denoiser a continuous score that says how badly the input image is degraded. It claims that this score, regressed from the PSNR between a degraded image and its reference, lets the same network handle diverse water conditions by rescaling the noise schedule and modulating feature normalization through a module called AdaGN. If true, the approach would improve on prior state of the art on UIEB and LSUI, raising PSNR to 28.60 dB and 28.57 dB respectively while also improving no-reference quality scores. A sympathetic reading is that degradation-aware conditioning, not just a bigger diffusion backbone, is what carries the improvement.","feed_headline":"Degradation-aware diffusion model tops underwater image benchmarks","feed_subtitle":"A predicted degradation score steers each denoising step, lifting PSNR to 28.60 dB on UIEB and 28.57 dB on LSUI.","key_machinery":"The load-bearing object is the degradation score D, a scalar in [0,1] computed as 1 minus the normalized PSNR between a degraded underwater image and its clean reference. D is injected into a conditional DDPM in two places: it rescales the noise schedule via beta_t * (1 + alpha D), and it drives a two-layer MLP that produces scale and shift parameters for Adaptive Group Normalisation (AdaGN), a variant of GroupNorm in which the per-group mean/variance normalization is modulated by the conditioning signal. The denoiser is a Swin-UNet backbone, and a Physical-guided Fusion Module (PGFM) amplifies the red channel by a degradation-dependent factor and applies frequency-domain attention. Together these components make the denoising trajectory depend on how degraded the input is, which is the mechanism the paper credits for adaptive enhancement.","core_discovery":"The central claim is that DACA-Net, by conditioning a DDPM on an estimated degradation score D, achieves state-of-the-art underwater image enhancement. The score D is a normalized PSNR-derived scalar in [0,1] regressed from paired raw and reference images by a dual-stream CNN; it enters the denoising network twice, as a multiplier on the noise variance beta_t and as the input to an Adaptive Group Normalisation (AdaGN) layer that shifts and scales feature statistics. A Physical-guided Fusion Module adds red-channel amplification and frequency-domain attention. On UIEB the method reports PSNR 28.60 dB / SSIM 0.9456 and on LSUI 28.57 dB / 0.9344, ahead of the compared methods including GUPDM and DM_underwater; ablations attribute the gain to the joint action of AdaGN, PGFM, and the hybrid histogram-perceptual-contrastive loss.","pith_inferences":["The reference-dependent definition of D (Eq. 1) means the training-time and test-time inputs of the estimator differ; a natural extension the paper does not spell out is a fully no-reference degradation regressor, which would make the method applicable to unpaired real footage.","The adaptive noise schedule beta_t * (1 + alpha D) could be tested in isolation: if alpha is learned to be near zero on some datasets, the claimed adaptivity is not actually engaged on those data.","The same degradation-conditioning recipe is transferable to other restoration domains with a scalar severity label, such as haze, rain, or low light, and the paper's ablations suggest the score, not the backbone, is the main driver.","Comparing DACA-Net against itself with D set to a constant would directly quantify how much of the 28.60 dB comes from degradation awareness rather than from the Swin-UNet and hybrid loss."],"forward_implications":["On the paper's evidence, replacing a fixed diffusion noise schedule with one scaled by a predicted degradation score improves PSNR and SSIM on UIEB and LSUI over fixed-schedule diffusion baselines.","The same conditioning scheme carries over to new datasets without retraining the score estimator if the PSNR range for normalization is known or estimated.","The degradation estimator is cheap enough (1.75 ms, 0.27 GFLOPs) that the adaptive conditioning adds little latency, so the framework is deployable near real time on a single GPU.","Because D is a continuous PSNR-based label, the network's behavior can be probed by construction: feeding a higher D should force stronger denoising and more aggressive red-channel compensation."],"supporting_citations":[{"why":"Bases the forward and reverse DDPM formulation and the noise schedule that the degradation score rescales.","marker":"[20]"},{"why":"Supplies the UIEB dataset and reference-pair benchmark used for training and evaluation.","marker":"[34]"},{"why":"Supplies the LSUI dataset used as the second benchmark for measuring generalization.","marker":"[45]"},{"why":"A diffusion-based underwater enhancement baseline whose fixed-schedule approach DACA-Net claims to improve on.","marker":"[56]"},{"why":"The previous state of the art on the two benchmarks, the main comparison to beat.","marker":"[42]"},{"why":"The group normalization scheme that AdaGN extends by conditioning scale and shift parameters on the degradation score.","marker":"[60]"},{"why":"The Swin-UNet backbone chosen for the denoiser, providing multi-scale hierarchical features.","marker":"[6]"}],"fun_headline_variants":["Degradation score tunes diffusion for sharper underwater images","DACA-Net: degradation-aware diffusion beats SOTA underwater","Adaptive noise from predicted degradation lifts PSNR to 28.6","Underwater clarity via score-conditioned denoising steps","DACA-Net: dual-stream predicts, diffusion restores, wins benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The degradation score that steers the whole enhancement is defined as a function of both the degraded image and its clean reference, and the paper does not describe how to compute it from a single raw image at test time.","fun_headline_variants_meta":{"raw":{"variants":["Degradation score tunes diffusion for sharper underwater images","DACA-Net: degradation-aware diffusion beats SOTA underwater","Adaptive noise from predicted degradation lifts PSNR to 28.6","Underwater clarity via score-conditioned denoising steps","DACA-Net: dual-stream predicts, diffusion restores, wins benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00018,"raw_usage":{"total_tokens":1310,"prompt_tokens":959,"completion_tokens":351,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":575,"completion_tokens_details":{"reasoning_tokens":263}},"tokens_in":575,"tokens_out":351,"duration_ms":5206,"temperature":1.0,"reasoning_tokens":263,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:36:10.134147+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained DACA-Net on a held-out underwater image with the reference withheld and trace whether the degradation estimator can produce D without the reference; if Eq. (1) is used as written, the conditioning signal is unavailable at inference and the claimed adaptive mechanism cannot operate. Alternatively, ablate D to a constant during denoising: if PSNR does not drop measurably, the degradation score is not carrying the improvement.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the LSUI dataset used as the second benchmark for measuring generalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"A diffusion-based underwater enhancement baseline whose fixed-schedule approach DACA-Net claims to improve on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The previous state of the art on the two benchmarks, the main comparison to beat."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The group normalization scheme that AdaGN extends by conditioning scale and shift parameters on the degradation score."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The Swin-UNet backbone chosen for the denoiser, providing multi-scale hierarchical features."}],"review_version":1}