{"id":"ac7533fb-a0bc-47ad-8791-1b1f44f43211","arxiv_id":"2507.06234","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A CLIP prompt-based underwater quality model, used as a training loss and curriculum weighting signal, improves NU2Net's U90 PSNR by 0.45 dB.","lead":"This paper adds a CLIP-based perception score and a curriculum contrastive loss to an underwater image enhancement network, reporting small but consistent gains over prior methods on a benchmark. It is worth reading as a case study in repurposing a general vision-language model as a domain-specific image quality metric.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table III contradicts the generalization claim: the proposed method's CLIP-Score is below NU2Net on U45 and C60, and UIQM below on SQUID and C60.","rationale":"The reader's weakest assumption concerned the CLIP perception model's generalization to U45, SQUID, and C60. My stress test goes further: even granting that assumption, the paper's own Table III does not support the generalization claim. The proposed method's CLIP-Score is worse than NU2Net's on two of the three non-reference datasets, and its UIQM is worse on two of three. The only metric where it consistently wins is UCIQE, which the paper itself dismisses as biased. This internal inconsistency is more concrete than the generalization-assumption concern and directly targets the abstract's headline claim. The U90 full-reference evidence (PSNR 23.115 vs 22.669) does support the core contribution of the two loss modules, so I do not recommend rejection, but the paper must either correct the generalization claim or provide independent perceptual validation. Since the reader already returned CONDITIONAL, I see no change to the overall verdict, only a sharper specification of the condition.","tokens_in":19289,"tokens_out":6275,"duration_ms":60824,"concrete_test":"Run a paired human preference study on U45, SQUID, and C60 comparing the proposed method's outputs with NU2Net's outputs, using at least 20 raters and forced-choice comparisons on all images. If the proposed method is not preferred over NU2Net at a statistically significant rate on at least two of the three datasets, the claimed generalization advantage is unsupported and the abstract should be revised.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract claims the method \"outperforms state-of-the-art methods in terms of visual quality and generalization ability,\" but the paper's own non-reference results in Table III contradict this. On U45, the proposed method's CLIP-Score (56.02) is lower than NU2Net's (56.89); on C60, it is 50.41 vs NU2Net's 50.81. On UIQM, the method is below NU2Net on SQUID (2.360 vs 2.480) and C60 (2.810 vs 2.900). The method's only consistent advantage is UCIQE (e.g., 0.601 vs 0.595 on U45), a metric the paper itself criticizes in Section IV.B as biased. Because the CLIP-Score is the very quantity optimized by Lclip and by the weighting in Eq. (6), reporting it as evidence of generalization is circular, and the numbers show no consistent superiority even on that optimized metric. Thus the generalization claim is neither independently validated nor internally consistent.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a plug-and-play perceptual training scheme for underwater image enhancement (UIE). The authors first train a CLIP-based perception model on the UEQAB MOS dataset using learnable antonymic prompt pairs (Section III.B, Eqs. 1-4). This model is then used in two ways inside the enhancement network: as a CLIP perception loss Lclip (Eq. 5) and as a difficulty-weighting criterion for curriculum contrastive regularization LCR (Eqs. 6-7), where negatives come from fixed UIE methods (UDCP, IBLA, DCP, HE, FUnIE, USUIR). The enhancement backbone is NU2Net, and the method is also applied to WaterNet, FUnIE, Shallow-UWnet, and PUGAN. Full-reference evaluation on U90 reports PSNR 23.115, SSIM 0.929, LPIPS 0.133 versus NU2Net's 22.669, 0.924, 0.154. No-reference evaluation is reported on U45, SQUID, and C60 using UCIQE, UIQM, and a CLIP-Score, plus qualitative comparisons and a saliency-detection application test.","tokens_in":19581,"tokens_out":5570,"duration_ms":57728,"significance":"If the central claim is read narrowly, the paper provides a consistent ablation story: adding Lclip and LCR to NU2Net improves full-reference metrics on U90 (Table IV), and adding both losses to four other UIE backbones improves them as well (Table VIII). The CLIP perception model itself achieves competitive PLCC/SROCC on the UEQAB test set (Table I), and the application test with a saliency detector is a useful auxiliary demonstration. The main weaknesses are that the headline full-reference gains are small and are reported without any variance or significance analysis, and the no-reference evaluation is partly circular because the reported CLIP-Score is computed by the same model that is used as the optimization target and as the curriculum-weighting criterion. The abstract's generalization claim is not supported by Table III, which shows the proposed method below NU2Net on several no-reference metrics. These issues are fixable by narrowing claims, adding multi-seed statistics, and validating the CLIP model on the no-reference datasets, so the work is potentially publishable but needs substantive revision.","major_comments":[{"comment":"Reporting CLIP-Score as a no-reference evaluation metric is circular for this method. The CLIP perception model that computes CLIP-Score is the same model optimized by Lclip (Eq. 5) and used to set the curriculum weights in Eq. (6), and it is validated only on UEQAB in Table I; Section IV.J itself concedes that the model \"may benefit from a more accurate selection of initialization prompts.\" Moreover, Table III does not show consistent superiority on the no-reference datasets: the proposed method is below NU2Net in CLIP-Score on U45 (56.02 vs 56.89) and C60 (50.41 vs 50.81), and below NU2Net in UIQM on SQUID (2.360 vs 2.480) and C60 (2.810 vs 2.900). Therefore the abstract's claim that the method \"outperforms state-of-the-art methods in terms of visual quality and generalization ability\" is not supported. The authors should either add human MOS studies on U45/SQUID/C60, report an external no-reference metric that is not part of the training objective, or restrict the generalization claim to the full-reference U90 result.","section":"Section IV.A.3 and Table III"},{"comment":"The full-reference gains on U90 are small (PSNR +0.446 dB, SSIM +0.005, LPIPS -0.021 relative to NU2Net), and every number in the paper appears to come from a single training run. No standard deviation, confidence interval, paired per-image test, or training-seed information is reported anywhere, even though the network structure is identical to NU2Net and gains of a few tenths of a dB could plausibly arise from run-to-run variation. Please report mean and standard deviation over multiple seeds and, where possible, paired per-image differences on U90 to establish that the proposed loss terms are the cause of the observed improvement.","section":"Tables II, IV, and VIII"},{"comment":"The SQUID test protocol is underspecified: the text says \"we select 16 representative examples as the test set, which is the same as [16],\" but gives no criterion for selecting these 16 examples. Because the generalization claim depends on the no-reference datasets U45, SQUID, and C60, the selection procedure must be described in advance or the full SQUID set should be used; otherwise the result cannot be distinguished from favorable subset selection.","section":"Section IV.A.1"}],"minor_comments":[{"comment":"The text is internally inconsistent: it states \"the scheme L1 + Lclip actually means NU2Net method [20],\" but Table V shows that NU2Net's reported performance corresponds to L1 + LUranker (22.669), while L1 + Lclip gives 22.677 in the same table. The paragraph also names the best scheme as \"L1 + LUranker\" twice, with one instance presumably meant to be L1 + LBFEN. Please rewrite this paragraph and reconcile it with Table V.","section":"Section IV.E"},{"comment":"The text refers to \"The SUIQD dataset\" in the discussion of the no-reference datasets; this appears to be a typo for SQUID and should be corrected.","section":"Section IV.C"},{"comment":"There is a discrepancy in the number of negatives: Algorithm 2 sets z=6 non-easy negatives and Section IV.A.2 lists six fixed UIE methods, but Section IV.F says \"the number of negatives is set as 7\" and labels the scheme \"+CR(1:7).\" Please clarify whether z excludes or includes the easy negative and align the notation.","section":"Algorithm 2 and Section IV.F"},{"comment":"The caption states that the top three results are marked with red, blue, and green, but the text does not explain or discuss these color markers, and the printed table does not make the ranking visually clear. Either explain the color coding in the text or remove it.","section":"Table III caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's most defensible evidence is the full-reference U90 ablation; the no-reference section should not be used to support the headline claim in its current form. The authors should also be encouraged to release code and trained models, since several claims depend on exact reproduction of the CLIP prompt-learning procedure and the fixed-negative generation pipeline. I do not see grounds for rejection if the generalization claim is narrowed and the statistical robustness of the small full-reference gains is demonstrated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers a genuine but modest result: adding a prompt-learned CLIP perception loss and CLIP-weighted curriculum contrastive regularization improves UIE networks on full-reference metrics. The plug-and-play experiments across four baselines are the strongest part. The generalization claim, however, is not supported by the paper's own no-reference table.\n\nWhat is actually new is the combination, not the ingredients. The antonymic-prompt CLIP quality estimator follows Wang et al. [21], and the curriculum contrastive regularization follows Zheng et al. [22]; the underwater setting and the use of the CLIP model to score hard vs. very-hard negatives are a legitimate extension. The paper does this cleanly. Table IV shows each loss module contributes, and Table VIII shows consistent PSNR/SSIM gains when the proposed losses are added to WaterNet, FUnIE, Shallow-UWnet, PUGAN, and NU2Net. That is real evidence and the best part of the paper.\n\nThe soft spots are real but not fatal to the core claim. First, the abstract and conclusion claim superior generalization, but Table III contradicts that: on U45 and C60 the method's CLIP-Score is below NU2Net, and UIQM is below NU2Net on SQUID and C60. The method only leads on UCIQE, a metric the paper itself criticizes as biased. So the generalization claim is not internally consistent. Second, the CLIP-Score is produced by the very model being optimized (Eq. 5 and Eq. 6), making it circular as an evaluation metric; the authors should not lean on it. The full-reference metrics are external and do improve, so the central claim survives, but the no-reference evidence should be presented much more carefully. Third, the gains are small: 0.446 dB PSNR over NU2Net on U90, with no variance or significance test, so one has to wonder about run-to-run noise. Fourth, the SQUID test selection of \"16 representative examples\" has no stated criteria. Fifth, there are textual errors: Section V.E's comparison is garbled, and the statement that \"L1 + Lclip actually means NU2Net\" is confusing and probably wrong. No code is released, which hurts reproducibility.\n\nThe limitation section is honest about prompt initialization sensitivity, which helps the paper's credibility.\n\nWho is this for? Researchers working specifically on underwater image enhancement who want a plug-and-play perceptual constraint. The full-reference gains are consistent across baselines, so the method is likely to be useful. But the paper overstates its generalization, and the evidence quality is weaker than the presentation suggests. It deserves a serious referee, not a desk reject, but a major revision is needed: tone down the generalization claim, report variance, fix the circular evaluation, clarify the textual errors, and release code.\n\nRecommendation: send to peer review, with the expectation of heavy revision.","headline":"Solid but modest plug-and-play CLIP losses improve UIE networks on full-reference metrics; the paper's generalization claim is contradicted by its own no-reference numbers.","tokens_in":20062,"tokens_out":2475,"would_cite":false,"duration_ms":25515,"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":"Adding a CLIP perception model trained on human opinion scores as a loss and curriculum regularizer improves underwater enhancement quality and generalization over state-of-the-art methods.","keywords":["underwater image enhancement","CLIP","perception loss","curriculum contrastive regularization","prompt learning","image quality assessment","no-reference evaluation","vision-language models"],"falsifier":"Train the same NU2Net-based UIE network with L1 + L_clip + L_CR but replace the CLIP score used to classify hard/very-hard negatives with random labels (or with the negative index order), keeping everything else identical. If the PSNR/SSIM/LPIPS results on U90 do not degrade materially compared to the proposed weighting, then the CLIP-based curriculum, not the mere presence of extra negatives, is not the source of the gains. Additionally, run a human preference study on 50 randomly sampled enhanced images from U45 and SQUID comparing the proposed method against NU2Net; if human raters do not choose the proposed method more often, the central perceptual-quality claim lacks support.","tokens_in":19141,"feed_emoji":"🌊","tokens_out":7211,"duration_ms":67023,"temperature":0.7,"pith_summary":"The paper argues that underwater image enhancement can be substantially improved by injecting human-perceptual judgment into the training loop through a CLIP vision-language model. It trains a small CLIP-based perception model—a softmax over cosine similarities between the image and two learnable prompts, 'Clear Underwater photo.' and 'Turbid Underwater photo.'—against mean opinion scores, then uses that model twice: as a perception loss that pushes enhanced images to look better than the reference, and as a scoring criterion that divides degraded 'negative' images into hard and very-hard classes for a curriculum contrastive regularizer. On the U90 test set the resulting network exceeds the NU2Net baseline it is built on (PSNR 23.115 vs 22.669, SSIM 0.929 vs 0.924, LPIPS 0.133 vs 0.154). The authors further claim the gains transfer to no-reference datasets U45, SQUID, and C60, and that the two loss modules can be added to other UIE networks plug-and-play without changing their structure. A careful reader should care because the method targets the widely noted perceptual shortcomings of existing UIE networks, not just pixel-level accuracy.","feed_headline":"CLIP perception loss beats underwater enhancement baselines","feed_subtitle":"A prompt-learned CLIP quality score, used as a loss and regularizer, lifts PSNR and perceptual fidelity on underwater test sets.","key_machinery":"The load-bearing object is the CLIP perception model: a frozen CLIP image encoder and text encoder with a learnable antonymic prompt pair (tokens T_p and T_n, both in $R^{{Nx512}}$), whose softmax of cosine similarities S_out = $e^{{S_p}}$/($e^{{S_p}}$+$e^{{S_n}}$) yields a predicted quality score in [0,1]. It is trained by L2 regression to human MOS labels on UEQAB (PLCC 0.83, SROCC 0.80 on the test set). Once trained, the prompt embeddings are frozen, and the model serves two roles: (a) it defines L_clip, a margin-style loss that drives the enhanced image's score above $\\alpha$ times the reference's deficiency; and (b) it evaluates each of the six non-easy negatives N_q at each epoch, classifying each as 'hard' (if S_out^Y' > S_out^Nq, weight 1+gamma) or 'very-hard' (otherwise, weight 1-gamma), with gamma=0.25. The L_CR term then computes a weighted ratio of VGG-19 feature L1 distances between positive/negative/anchor, pulling the anchor to the reference and pushing it from a weighted combination of degraded images.","core_discovery":"The central discovery is that a CLIP perception model trained on human opinions can act as a differentiable quality oracle for underwater images, and that using it simultaneously as a loss and as a curriculum weighting of negatives yields better enhancement than the same network trained with L1 alone or with earlier quality-assessment losses. Concretely, the paper's L_clip = max(0, (1-S_out(Y')) - alpha(1-S_out(Y))) with alpha=0.975 encourages the enhanced image Y' to achieve a CLIP score exceeding the reference, while L_CR contrasts VGG-19 features of the anchor Y' against the reference Y, the 6 non-easy negatives from fixed UIE methods, and the input; the weight of each non-easy negative is 1+gamma or 1-gamma depending on whether its CLIP score is below or above the current anchor's, making difficult negatives gradually easier over training. The ablation tables show each module helps, and that the CLIP-based version of L_CR outperforms the PSNR-based version, which is the load-bearing evidence for using CLIP as the quality criterion.","pith_inferences":["Since the paper validates the CLIP perception model's correlation with human scores only on UEQAB, a natural test is to collect human MOS on the U45, SQUID, and C60 enhancement outputs; if CLIP scores disagree with human rankings there, the reported no-reference CLIP-Score improvements should be interpreted as model preference rather than human-perceived quality.","The binary hard/very-hard split with fixed gamma could be replaced by a continuous weight based on the CLIP score difference, which might yield smoother curriculum dynamics and faster convergence.","The six negative-generation methods are fixed and heuristic; using the CLIP perception model to also select or generate negatives (e.g., through learned degradation models) could strengthen the regularization and push the gain further.","The plug-and-play claim suggests a drop-in use for other underwater tasks like depth estimation or saliency detection, a direction the paper gestures at only through the SVAM-Net saliency experiment."],"forward_implications":["Any existing UIE network (WaterNet, FUnIE, Shallow-UWnet, PUGAN, NU2Net) gains PSNR and SSIM when L_clip and L_CR are added, so the two loss modules are architecture-agnostic additions.","The CLIP perception model can serve as a no-reference evaluation score for underwater images, complementing UCIQE and UIQM, since it matches human MOS more closely on UEQAB.","Training with content-matched negatives at varying difficulty levels constrains the solution space and prevents both under- and over-enhancement, as shown by the ablation with single-negative and nine-negative configurations.","The method generalizes to no-reference test sets with diverse distortion types (color casts, low contrast, blur) without retraining on those sets.","Because L_clip is differentiable, the same perception model could be reused for other restoration tasks that optimize human-perceived quality."],"supporting_citations":[{"why":"Supplies the antonymic prompt pair method for using CLIP to assess perceptual image quality.","marker":"[21]"},{"why":"The NU2Net network and URanker loss that the proposed method builds on and compares against as the strongest baseline.","marker":"[20]"},{"why":"Introduces curricular contrastive regularization with non-easy negatives for dehazing, which the paper adapts to underwater enhancement.","marker":"[22]"},{"why":"Provides the UEQAB dataset of 8000 MOS-labeled underwater images used to train the CLIP perception model, plus the UIQAN integration idea.","marker":"[49]"},{"why":"The pretrained CLIP model that supplies the image and text encoders.","marker":"[23]"},{"why":"The UIEB benchmark dataset and its train/test split that provide the reference images and the C60 challenging set.","marker":"[18]"},{"why":"The LPIPS perceptual metric used to evaluate enhancement fidelity against references.","marker":"[51]"}],"fun_headline_variants":["CLIP perception loss boosts underwater image quality","Underwater enhancement with CLIP-guided curriculum","CLIP as a quality oracle for underwater images","Curriculum contrastive learning meets CLIP for underwater","CLIP loss sharpens underwater image enhancement"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The CLIP perception model trained on UEQAB mean opinion scores accurately judges perceptual quality on all the other test sets (U45, SQUID, C60), so that both the L_clip loss and the weighting of negatives in L_CR are guided by a trustworthy human-aligned quality score.","fun_headline_variants_meta":{"raw":{"variants":["CLIP perception loss boosts underwater image quality","Underwater enhancement with CLIP-guided curriculum","CLIP as a quality oracle for underwater images","Curriculum contrastive learning meets CLIP for underwater","CLIP loss sharpens underwater image enhancement"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000686,"raw_usage":{"total_tokens":3164,"prompt_tokens":1054,"completion_tokens":2110,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":2042}},"tokens_in":670,"tokens_out":2110,"duration_ms":16301,"temperature":1.0,"reasoning_tokens":2042,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T05:30:13.008648+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same NU2Net-based UIE network with L1 + L_clip + L_CR but replace the CLIP score used to classify hard/very-hard negatives with random labels (or with the negative index order), keeping everything else identical. If the PSNR/SSIM/LPIPS results on U90 do not degrade materially compared to the proposed weighting, then the CLIP-based curriculum, not the mere presence of extra negatives, is not the source of the gains. Additionally, run a human preference study on 50 randomly sampled enhanced images from U45 and SQUID comparing the proposed method against NU2Net; if human raters do not choose the proposed method more often, the central perceptual-quality claim lacks support.","supporting_citations":[{"cited_title":"Underwater ranker: Learn which is better and how to be better,","cited_arxiv_id":null,"evidence_quote":"The NU2Net network and URanker loss that the proposed method builds on and compares against as the strongest baseline."},{"cited_title":"Curricular contrastive regularization for physics-aware single image dehazing,","cited_arxiv_id":null,"evidence_quote":"Introduces curricular contrastive regularization with non-easy negatives for dehazing, which the paper adapts to underwater enhancement."},{"cited_title":"Human perceptual quality driven underwater image enhancement framework,","cited_arxiv_id":null,"evidence_quote":"Provides the UEQAB dataset of 8000 MOS-labeled underwater images used to train the CLIP perception model, plus the UIQAN integration idea."}],"review_version":1}