{"id":"299ec4e1-24e6-4250-bfb6-8a6c19275e3c","arxiv_id":"2412.07693","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A CLIP-guided training recipe for zero-reference low-light enhancement improves downstream detection and classification on several benchmarks, though gains are small and the evaluation has fairness gaps.","lead":"This paper adds two CLIP-based training signals, learned quality prompts and content/context text guidance, to a zero-reference low-light image enhancer. The aim is to make dark-scene photos easier for object detectors and classifiers to read, and the paper reports small gains on several low-light benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central empirical claim rests on task-based gains, but the paper never states a train/test split, and three of the five evaluation datasets are in the training set; this is the load-bearing validity threat.","rationale":"The strongest claim is that adding a CLIP prior and content/context semantic guidance yields consistent task-based gains over zero-reference methods. For that claim to hold, the evaluation must measure generalization to low-light images not used in training. The manuscript never gives a split, and the training set contains three of the four evaluation datasets (NOD, ExDark, DarkFace) plus ExLPose, LOL, and BAID. This makes the reported mAP and accuracy gains uninterpretable. I checked alternative concerns: the Table VII comparison uses original checkpoints for paired/unpaired methods while \"Ours\" is trained on the target datasets; the full model underperforms its own baseline on ExDark classification accuracy in Table III; and no variance estimates are reported. These are real but secondary: Table IV does retrain zero-reference competitors on the same training set, and the full model is still best or second-best on most rows. The train/eval overlap issue cuts at the central claim and is the most load-bearing. If a disjoint split is provided, the empirical claim may survive; as written, it cannot be verified. Therefore I agree with the reader's REJECT and recommend no change.","tokens_in":22460,"tokens_out":3824,"duration_ms":36519,"concrete_test":"Obtain the exact training and evaluation image lists from the released repository (github.com/igor-morawski/tmm-sem). Compute exact or perceptual hashes for every training image/quadrant and every evaluation image in NOD, NOD SE, ExDark, and DarkFace. Exclude all evaluation images that appear (or near-duplicate) in the training set, or better, retrain using official train splits and evaluate on official test splits. If the task-based improvements over Zero-DCE and SCI shrink or reverse on the disjoint subset, the paper's central generalization claim is unsupported; if the improvements persist, the leakage concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A states the enhancement model is trained on \"a collection of data extracted from the NOD, ExDark, DarkFace, ExLPose, LOL and BAID datasets.\" Section IV-D then evaluates on NOD, NOD SE, ExDark, and DarkFace (Tables III, IV, VII) with no stated train/test split. Three of the four evaluation benchmarks are thus in-distribution unless an explicit split is provided. Additionally, the CLIP projection layer is fine-tuned on content/context descriptions from these datasets, so the model can exploit dataset-specific cues rather than learning a general enhancement operator. Because the central claim is a \"consistent\" task-based improvement across many low-light datasets, in-sample evaluation would invalidate the comparison: the gains in Tables III, IV, and VII could reflect memorization, not enhancement quality. The paper's acknowledged limitations (need for annotation, weak denoising) do not address this evaluation gap. A held-out split or released training/evaluation lists is required before the empirical claim can be assessed.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a zero-reference low-light image enhancement method that builds on Zero-DCE by adding a learned CLIP-based image prior (via prompt learning with an averaging/subsampling augmentation) and two semantic guidance losses based on abstracted content and context descriptions derived from existing bounding-box annotations. The enhancement model is trained without paired or unpaired normal-light data, and the method is evaluated by downstream task performance (object detection, face detection, image classification) on NOD, NOD SE, LOD, ExDark, and DarkFace, as well as by PSNR/SSIM on VE-LOL. The central claim is that the proposed prior and semantic guidance lead to consistent task-based improvements over related zero-reference methods.","tokens_in":22672,"tokens_out":6540,"duration_ms":53720,"significance":"If the claims hold, the work would provide a lightweight, training-time-only enhancement module that improves machine cognition in low light without requiring paired or unpaired normal-light data, while leveraging open-vocabulary CLIP capabilities to avoid fixing the annotated category set. The paper's focus on task-based evaluation with external detectors (YOLO-World, YOLO5Face) and its explicit discussion of the non-correlation between restoration metrics and downstream task performance are strengths, as is the inclusion of an ablation study and a generalization study across multiple zero-reference baselines. However, the central empirical claim is currently compromised by the lack of a stated train/evaluation split and by several internal inconsistencies in the reported results.","major_comments":[{"comment":"The enhancement model is trained on images extracted from NOD, ExDark, DarkFace, ExLPose, LOL, and BAID (Section IV-A), and the CLIP projection layer is fine-tuned on content/context descriptions from these data. The task-based evaluation in Tables III, IV, and VII then reports results on NOD, NOD SE, ExDark, and DarkFace without specifying any train/test split. Because three of the four evaluation benchmarks are part of the training corpus, the reported gains over the no-enhancement baseline and over methods evaluated with their original checkpoints may reflect memorization rather than a general enhancement operator. The authors must state the exact split, release training/evaluation lists, or re-run the comparisons on held-out subsets of these datasets and on datasets not used for training before the central claim can be assessed.","section":"IV-A and IV-D"},{"comment":"The label assignment in the prompt-learning loss appears inverted relative to the intended semantics. The text states that y = 0 for a positive (averaged) image and y = 1 for a negative (subsampled) image, while Eq. (2) defines ŷ as the softmax probability that an image matches the positive prompt. With the stated assignment, the binary cross-entropy loss pushes the averaged image toward the negative prompt and the subsampled image toward the positive prompt, which is the opposite of the described 'positive quality prompt' and 'negative quality prompt' behavior. Please clarify whether this is a typographical error in the text or an actual property of the implementation; if the implementation follows the equations as written, the learned prior would be the inverse of what is claimed and the qualitative benefits attributed to it would be unexplained.","section":"III-A, Eqs. (1)-(2)"},{"comment":"The comparison against paired and unpaired methods is inequitable because the competitors (SNR+SKF, DRBN+SKF, PairLIE, DiffLL, CLIP-LIT, PIE, QuadPrior) are evaluated using their original public checkpoints, whereas Ours is trained on the evaluation datasets themselves (NOD, ExDark, DarkFace, and others). The paper's statement that a cross-dataset setup was also reported in recent works does not remedy the asymmetry when the proposed method is trained on the target benchmarks. This setup conflates the effect of the proposed enhancement losses with the advantage of in-distribution training, so the rankings in Table VII cannot be interpreted as evidence of the method's generalization.","section":"IV-D, Table VII"},{"comment":"The claim of 'consistent improvements' over the baseline is not supported by the reported numbers. In Table IV, Ours achieves 42.5% classification accuracy on ExDark, below the Zero-DCE baseline's 44.1%; in Table VII, PairLIE reaches 47.7% on the same metric, while Ours obtains 42.5%. Moreover, Table IX shows that applying the proposed strategy to SCI decreases mAP on NOD (43.9% to 43.7%), NOD SE (27.0% to 26.7%), and LOD (45.6% to 45.4%), which contradicts the sentence in Section IV-G that states 'we observe significant improvement in all methods with exception of SCI.' The authors should revise the consistency claim and explicitly discuss these exceptions.","section":"IV-B and IV-D, Tables IV and IX"}],"minor_comments":[{"comment":"The exposure loss is written with an undefined double sum '1/M \\sum_{i=1}^M \\sum_{i=1}^X |\\hat{I}_i - E|'; it should presumably be a single sum over the M non-overlapping 16x16 patches, with the index consistent throughout the equation.","section":"III-B, Eq. (5)"},{"comment":"Equation (10) has a missing closing parenthesis in 'cos(Φ_img(Î_i), Φ_txt(A_j)' and should read 'cos(Φ_img(Î_i), Φ_txt(A_j))'.","section":"III-B, Eq. (10)"},{"comment":"The phrase 'with any limitations on annotated object categories' appears to be a typo and should read 'without any limitations on annotated object categories'.","section":"Fig. 3 caption"},{"comment":"The sentence 'Focusing on leveraging the CLIP [17] model for semantic segmentation' is confusing, since the paper does not perform semantic segmentation; please rephrase to describe the actual use of CLIP for semantic guidance.","section":"III-B, first paragraph"},{"comment":"The abbreviation 'NOD SE' is used in the tables and text but is never defined; please define it in Section IV (it appears to refer to a subset or extreme-light variant of NOD).","section":"IV-D and Tables III-VII"}],"recommendation":"major_revision","confidential_remarks":"The data-leakage issue in Section IV is the most serious concern and alone would justify rejection if the authors cannot provide a proper train/test split or demonstrate held-out performance. The inverted label assignment in Section III-A is also potentially fatal; if the implementation matches the equations, the prompt-learning stage would train the opposite of the intended association, and the reported qualitative improvements would be inexplicable. I would urge the editor to request the authors to release exact data splits and to verify the prompt-learning labels in the released code before any further review cycle. The paper's self-citation pattern (building directly on the authors' prior CVPR 2024 work and using their NOD dataset as a benchmark) is not disqualifying but should be kept in mind when assessing novelty."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the content/context batch-matching idea is a real, if modest, extension of the authors' CVPR 2024 work, and the resampling-based prompt prior is a neat trick. The paper is clearly written and the ablations are organized. But the central numerical claim — consistent task-based gains over zero-reference baselines — is not yet supported because the evaluation protocol is not disclosed. Section IV-A trains on NOD, ExDark, DarkFace, ExLPose, LOL, and BAID; Section IV-D evaluates on NOD, NOD SE, ExDark, and DarkFace with no stated split. Three of those four are in the training mix. Unless the authors release train/test lists, the gaps in Tables III, IV, and VII could be in-distribution memorization rather than enhancement quality. This is not a nitpick; it is the load-bearing assumption behind every table in the paper.\n\nWhat is genuinely new: the two-step content/context guidance goes beyond the single-instance label classification in [16], and the paper honestly tests against external detectors (YOLO-World, YOLO5Face). The qualitative study and the observation that restoration metrics do not transfer to task gains are useful.\n\nSoft spots other than the split: the full model underperforms its own Zero-DCE baseline on ExDark classification accuracy (42.5 vs 44.1 in Table IV), and the Table VII comparison pits original-checkpoint competitors against a model trained on target datasets. No variance estimates anywhere. The authors do disclose limitations (need for annotation, weak denoising) but those are separate from the evaluation gap. Also, the paper does not overclaim: the gains are 1-4 mAP points, and they say so.\n\nBottom line: this is a serious paper that deserves referee time, but the evaluation has to be fixed — either by holding out all evaluation datasets or by releasing precise split information. Without that, the paper is not publishable in its current form. If the authors can provide splits, the result would be a reasonable journal contribution: a lightweight, zero-reference enhancement module with semantic guidance.","headline":"Useful extension of the authors' CLIP-guided zero-reference enhancement, but the missing train/test split makes the headline gains unsubstantiated.","tokens_in":23215,"tokens_out":1709,"would_cite":false,"duration_ms":15257,"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 CLIP-learned image priors and content/context semantic guidance to a zero-reference enhancer improves machine cognition in low light without paired training data.","keywords":["low light","low-light enhancement","semantic guidance","unsupervised enhancement","CLIP","prompt learning","zero-reference enhancement","machine cognition"],"falsifier":"Train the full method on a strict disjoint split of the same datasets (e.g., train only on explicitly held-out partitions of NOD, ExDark, and DarkFace, and evaluate on the remaining images from those datasets), then compare mAP and classification accuracy with the paper's tables; if the gains over the no-enhancement baseline vanish or shrink sharply, the central claim of generalization fails.","tokens_in":22237,"feed_emoji":"🌙","tokens_out":7310,"duration_ms":56434,"temperature":0.7,"pith_summary":"Low-light images confuse computer vision systems, and collecting paired normal-light training data is expensive. This paper claims that a zero-reference enhancer—one that needs no paired or unpaired normal-light data—can be made better for machine cognition by adding two kinds of CLIP-based guidance: a learned image prior and semantic content/context cues. The image prior is learned by contrasting noise-averaged and noise-preserving resampled versions of low-light images, producing prompt embeddings that constrain contrast and noise. The semantic guidance converts existing box annotations into text descriptions of objects inside and outside each training patch, and trains the enhancer to match enhanced patches to those descriptions. Across object detection, face detection, and classification benchmarks, the paper reports consistent task-based gains over related zero-reference methods, along with reduced over-saturation and noise over-amplification.","feed_headline":"CLIP cues help a zero-reference enhancer see better in the dark","feed_subtitle":"No paired normal-light data needed; the same module lifts detection, classification, and face detection.","key_machinery":"The load-bearing mechanism is the joint training loss on a lightweight curve-estimation network, combining (1) the zero-reference losses of Zero-DCE (exposure, spatial consistency, color or Gray-World, and illumination smoothness), (2) a CLIP prompt-pair prior learned by contrasting averaged versus subsampled resamplings of the same low-light images, and (3) two cross-entropy semantic-guidance losses that match enhanced patches, via CLIP image/text embeddings, to abstracted comma-separated descriptions of objects inside the patch (content) and outside the patch (context). The content and context losses are computed on an $N \\times N$ similarity matrix in both directions, image-to-text and text-to-image. The prior constrains global contrast and noise amplification; the semantic guidance enriches color and foreground-background separability; together they modify what the curve parameters produce. The method uses Zero-DCE's pixel-wise curve application $L_{En}(x) = L_{En-1}(x) + A_n(x) L_{En-1}(x)(1-L_{En-1}(x))$ and keeps the baseline architecture unchanged.","core_discovery":"The paper's central claim is that injecting a CLIP-learned image prior and content/context semantic guidance into a zero-reference curve-based enhancement network (Zero-DCE) improves the network's output specifically for machine cognition, without any need for paired or unpaired normal-light training data. The learned prior is a pair of text prompts: a positive prompt trained to match m×m-averaged images (a fast proxy for denoising) and a negative prompt trained to match 1:m-subsampled images that retain noise, learned with cross-entropy loss. The semantic guidance is realized as two matching tasks within each training batch: the enhanced image patch must match text descriptions of the object instances inside the patch (content), and also descriptions of instances outside the patch (context). The paper reports that this combination raises mAP and classification accuracy on NOD, NOD SE, LOD, ExDark, and DarkFace relative to the Zero-DCE baseline and other zero-reference methods, while holding inference cost low, and it argues from the results that human-perceptual restoration quality (PSNR/SSIM) does not correlate with downstream task performance in low-light enhancement.","pith_inferences":["A testable extension the authors do not run: apply the content/context guidance to other degradations (dehazing, deraining, super-resolution) for machine-vision pipelines; if the semantic matching loss is the active ingredient, it should transfer to any restoration task with box or text annotations.","The learned prior's dependence on CLIP as a frozen encoder implies its ceiling is set by CLIP's ability to recognize objects in heavily degraded patches; using a vision-language model trained on low-light or noisy data could push the gains further, which the paper does not explore.","The authors' observation that QuadPrior over-smooths extreme low-light instances suggests a complementary direction: combining semantic guidance with an explicit denoising or inpainting module could recover instances that detection models currently miss, rather than relying on the zero-reference curve network which has no denoising capability."],"forward_implications":["A single enhancement module trained this way can be inserted in front of any downstream detector or classifier, improving low-light performance at negligible inference cost, since the semantic guidance and prompt prior act only at training time.","The training recipe requires only low-light images plus existing bounding-box annotations; for paired datasets without annotations, the authors auto-label with an open-vocabulary detector, so the method scales to new categories without fixing the training category set.","The same training strategy transfers to other zero-reference baselines (RUAS, SGZ, SCI, Zero-DCE), giving consistent task-based gains on all but SCI, where the authors hypothesize the extra losses interfere with its self-calibration module.","With 32× down/up-scaling of the curve parameter map at inference, the method retains most of its task-based gain while cutting computational complexity to 0.08 GMACs, roughly three orders of magnitude below the un-scaled baseline.","The reported decoupling of PSNR/SSIM from task accuracy supports the paper's claim that low-light enhancement for machine cognition should be evaluated by downstream task metrics, not by perceptual quality alone."],"supporting_citations":[{"why":"Supplies the baseline DCE-Net architecture and zero-reference losses that the method extends.","marker":"[13]"},{"why":"Provides the CLIP text and image encoders used for both the learned image prior and the semantic guidance.","marker":"[17]"},{"why":"The authors' earlier work that introduced unsupervised CLIP prompt prior and detection-instance semantic guidance, which this paper extends with content/context tasks.","marker":"[16]"},{"why":"Introduces prompt learning for an image prior using CLIP in a related unsupervised enhancement setting, the direct inspiration for the learned prior.","marker":"[28]"},{"why":"NOD dataset and its task-based evaluation protocol for object detection in extreme low light.","marker":"[24]"},{"why":"ExDark dataset used for training and evaluation of classification/detection in low light.","marker":"[1]"},{"why":"DarkFace dataset used for face-detection evaluation and training.","marker":"[20]"},{"why":"LOD dataset used to evaluate detection and to measure the impact of low-light level via α-blending.","marker":"[25]"},{"why":"YOLO-World open-vocabulary detector used for all object-detection evaluations.","marker":"[77]"},{"why":"YOLO5Face detector used for face-detection evaluations on DarkFace.","marker":"[80]"}],"fun_headline_variants":["CLIP prior and cues refine zero-reference low-light images","Zero-ref low-light enhancer gets CLIP boost without paired data","No paired data: CLIP prior and context cues lift zero-reference low-light","CLIP content and context cues sharpen low-light images for AI tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the images used for training (drawn from NOD, ExDark, DarkFace, ExLPose, LOL, and BAID) are disjoint from the images used for evaluation on NOD, NOD SE, ExDark, and DarkFace; the paper never states a train/test split, so if evaluation images overlap the training set, the reported gains could reflect memorization rather than generalization.","fun_headline_variants_meta":{"raw":{"variants":["CLIP prior and cues refine zero-reference low-light images","Zero-ref low-light enhancer gets CLIP boost without paired data","No paired data: CLIP prior and context cues lift zero-reference low-light","CLIP content and context cues sharpen low-light images for AI tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001307,"raw_usage":{"total_tokens":5372,"prompt_tokens":1033,"completion_tokens":4339,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":649,"completion_tokens_details":{"reasoning_tokens":4264}},"tokens_in":649,"tokens_out":4339,"duration_ms":28932,"temperature":1.0,"reasoning_tokens":4264,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T18:36:30.755825+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the full method on a strict disjoint split of the same datasets (e.g., train only on explicitly held-out partitions of NOD, ExDark, and DarkFace, and evaluate on the remaining images from those datasets), then compare mAP and classification accuracy with the paper's tables; if the gains over the no-enhancement baseline vanish or shrink sharply, the central claim of generalization fails.","supporting_citations":[{"cited_title":"Unsupervised image prior via prompt learning and clip semantic guidance for low-light image enhancement,","cited_arxiv_id":null,"evidence_quote":"The authors' earlier work that introduced unsupervised CLIP prompt prior and detection-instance semantic guidance, which this paper extends with content/context tasks."},{"cited_title":"NOD: Taking a Closer Look at Detection under Extreme Low-Light Conditions with Night Object Detection Dataset","cited_arxiv_id":"2110.10364","evidence_quote":"NOD dataset and its task-based evaluation protocol for object detection in extreme low light."},{"cited_title":"Advancing image understanding in poor visibility environments: A collective benchmark study,","cited_arxiv_id":null,"evidence_quote":"DarkFace dataset used for face-detection evaluation and training."},{"cited_title":"Crafting object detection in very low light,","cited_arxiv_id":null,"evidence_quote":"LOD dataset used to evaluate detection and to measure the impact of low-light level via α-blending."},{"cited_title":"Yolo5face: Why reinventing a face detector,","cited_arxiv_id":null,"evidence_quote":"YOLO5Face detector used for face-detection evaluations on DarkFace."}],"review_version":1}