{"id":"94955461-2246-4f22-a108-36f4c3ec8e33","arxiv_id":"2412.00890","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"CLAD, a contrastive vision-language model for industrial anomaly detection, reports higher scores than four baselines on MVTec-AD, but its own Table 1 shows it loses to WinCLIP on VisA pixel-level AUC.","lead":"This paper introduces CLAD, a method that uses contrastive learning to align product images with text descriptions for industrial defect detection, and reports results on MVTec-AD and VisA. The claimed advantage over existing methods is not supported by the paper's own tables, and the method is described too vaguely to reproduce.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 itself contradicts the headline claim: CLAD's VisA pixel-AUC (96.2) is below WinCLIP (96.4), so the paper's central 'highest scores on both datasets/tasks' claim is internally falsified.","rationale":"I read the paper in good faith as an attempt to show that a contrastive cross-modal alignment method, CLAD, outperforms existing anomaly detection methods on MVTec-AD and VisA. The central claim depends on the numerical results in Table 1 being accurate and supporting the stated superiority. The most load-bearing concern is therefore the internal contradiction in Table 1: CLAD's reported VisA Pixel-AUC is 96.2, which is lower than WinCLIP's 96.4. This directly conflicts with the claim of 'highest scores in both anomaly detection and localization tasks' and with the phrase 'consistently outperforms all other methods on both datasets.' This is not an ad hominem or a disagreement with the field's consensus; it is a simple check against the paper's own reported numbers. The reader's weakest_assumption pointed to the ill-defined training signal in Eq. 3, which is a real and substantial problem, and the reader's rationale also mentioned the Table 1 contradiction. My focus differs slightly because the Table 1 contradiction is the most direct and least assumption-dependent route to evaluating the central claim: even if every training detail were clarified, the claimed empirical superiority would still be false as stated unless the table is corrected. The concrete test I propose is deliberately narrow: comparing the two numbers in Table 1 settles whether this particular concern lands. If the authors provide a corrected table or an explanation that the 96.2 entry is a typo, then the central claim could potentially be re-evaluated. Without such a correction, the paper should not be accepted as is. Because this concern supports the reader's existing REJECT verdict rather than moving it, I recommend UNCHANGED.","tokens_in":9876,"tokens_out":2771,"duration_ms":25506,"concrete_test":"Re-read Table 1 and compare the VisA Pixel-AUC row for CLAD against the same row for WinCLIP. Verify numerically whether 96.2 is greater than 96.4. If it is not, the statement that CLAD achieves the highest scores in both tasks on both datasets is false as written. As a secondary check, confirm whether any other row in Table 1 also fails to support the 'consistently outperforms all other methods' claim when pairwise compared with WinCLIP.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, stated in Section 4.2 and the abstract, is that CLAD 'consistently outperforms all other methods on both datasets, achieving the highest scores in both anomaly detection and localization tasks.' For this claim to hold, Table 1 must show CLAD with the best Image-AUC and Pixel-AUC on both MVTec-AD and VisA. It does not. Under VisA Pixel-AUC, CLAD is reported as 96.2±0.1 while WinCLIP is reported as 96.4±0.4. With the numbers as printed, CLAD does not achieve the highest pixel-level AUC on VisA. This is not an external reproducibility question, a statistical subtlety, or a disagreement with published baselines; it is an internal contradiction between the paper's own evidence and its central claim. The claim is load-bearing because the entire contribution is defined by the reported superiority. If the table entry is a typo, the paper needs an erratum and a corrected comparison; if the table is accurate, the central claim should be weakened or the paper should be rejected as stated. The other concerns raised in review, such as the ill-defined contrastive training signal in Eq. 3 and the undefined inverse encoders in Eq. 7, are serious, but the Table 1 inconsistency is the most direct falsification of the paper's headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes CLAD, a method that uses a pretrained vision-language model to embed images and text, aligns them with a contrastive loss, and scores anomalies by cross-modal embedding distance, with Grad-CAM-style localization. Experiments on MVTec-AD and VisA report image-level and pixel-level AUC compared with SPADE, PaDiM, PatchCore, and WinCLIP, plus an ablation study, human evaluation, and IoU analysis. The paper claims CLAD achieves the highest scores in both detection and localization on both datasets.","tokens_in":10254,"tokens_out":5336,"duration_ms":47085,"significance":"The topic is timely, and a simple cross-modal alignment approach for industrial anomaly detection would be a useful contribution if the claimed results were supported. The manuscript, however, does not provide a working, reproducible specification of the method: the central experimental claim is contradicted by its own Table 1, the contrastive loss is ill-posed under the stated normal-only training setting, and the reconstruction loss depends on undefined inverse encoder functions. No code, no text prompts, no hyperparameters, and no evaluation protocols are given. The paper's strengths are its clear identification of the problem and the presence of ablations; but the evidence is insufficient to support the stated claims.","major_comments":[{"comment":"The central claim that CLAD \"consistently outperforms all other methods on both datasets, achieving the highest scores in both anomaly detection and localization tasks\" is directly contradicted by the numbers in Table 1: on VisA pixel-level AUC, CLAD is reported as 96.2±0.1 while WinCLIP is reported as 96.4±0.4. With the numbers as printed, CLAD does not achieve the highest pixel-level AUC on VisA. Since the entire contribution is defined by this reported superiority, the table must be corrected or the claim withdrawn.","section":"Section 4.2, Table 1"},{"comment":"The contrastive loss treats pairs (i,j), j≠i, as negative pairs. Section 4.1 describes the standard unsupervised setting with normal-only training images. Under normal-only training, those negative pairs are normal-normal pairs, so the loss pushes normal instances apart, which is the opposite of the stated goal of grouping normal instances. If labeled anomalous pairs are used instead, the method is supervised and no longer matches the claimed unsupervised IAD setting. The paper never resolves this ambiguity.","section":"Section 3.2, Eq. (3)"},{"comment":"The reconstruction loss uses inverse encoder functions f_v^{-1} and f_t^{-1}, which are never defined anywhere in the architecture or experimental setup. Pretrained CLIP-style encoders do not provide such inverses, and no decoder or inversion procedure is described. Consequently the total loss in Eq. (6) is not computable as stated, and the training procedure is incomplete.","section":"Section 3.4, Eq. (7)"},{"comment":"The human evaluation and IoU results are reported without any protocol. There is no information about the number of expert raters, the number of images rated, the exact rating instructions, whether the raters were blinded to method, or whether differences are statistically significant. Table 4 does not state on which dataset or split the IoU was computed, nor the threshold used to binarize the predicted anomaly maps. These results are load-bearing for the interpretability claim and cannot be evaluated as presented.","section":"Sections 4.4 and 4.5, Tables 3 and 4"},{"comment":"The baseline comparison protocol is underspecified. The paper does not state whether the SPADE, PaDiM, PatchCore, and WinCLIP numbers are taken from the original publications or from a reimplementation, nor does it report the backbone, input resolution, prompt templates, or threshold choices used for the baselines. Without this information, the comparison in Table 1 cannot be independently verified, even setting aside the internal contradiction noted above.","section":"Section 4.1 and Table 1"}],"minor_comments":[{"comment":"The introduction names AnomalyGPT, PaDiM, and PatchCore as prior approaches, but Table 1 does not include AnomalyGPT. Since AnomalyGPT is an LVLM-based anomaly detection method and CLAD is also LVLM-based, its omission weakens the comparison and should be justified.","section":"Section 1 and Section 4.1"},{"comment":"The symbols α_k and A_k in the Grad-CAM equation are not defined in the text; as written they are free symbols rather than the final-layer weights and activation maps of the visual encoder.","section":"Section 3.3, Eq. (5)"},{"comment":"The notation for the positive-part function is inconsistent: the first term uses [·]_+ while the second term uses max(0, ·). Use a single notation throughout.","section":"Section 3.2, Eq. (3)"},{"comment":"The anomaly threshold and the scaling factor σ are never specified, and no sensitivity analysis is provided. This makes the classification rule in Section 3.5 difficult to reproduce.","section":"Section 3.5 and Eq. (4)"},{"comment":"Reference [25] is cited in support of Vision Transformers for industrial image anomaly detection, but the cited paper is about exemplar convolutional neural networks, not vision transformers; the citation appears mismatched.","section":"Section 2.2, Ref. [25]"},{"comment":"The paper contains no limitations section. Given the claims of real-world applicability, the authors should discuss known limitations, such as reliance on textual descriptions and sensitivity to prompt design.","section":"Entire manuscript"}],"recommendation":"reject","confidential_remarks":"The reference list contains a notable number of self-citations to the first author's related work, several of which are not clearly relevant to the technical content (e.g., Refs. 9–11, 14–15, 18–19). I would suggest the editor check whether these citations are needed for the paper's argument or whether they are primarily self-promotional. The manuscript also lacks reproducibility artifacts, which is concerning for a results-heavy empirical claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, you should know two things about this one. First, the headline claim — that CLAD consistently achieves the highest image and pixel AUC on both MVTec-AD and VisA — is contradicted by the paper's own Table 1. On VisA pixel-level AUC, WinCLIP reports 96.4 and CLAD reports 96.2. That is not an external reproducibility quibble; it's an internal falsification of the central contribution. Second, the training loss in Eq. 3 is not coherent with the stated unsupervised setting, where only normal images are used. The negative pair term pushes different instances apart, which would separate normal samples from each other. If the authors used labeled anomalies at training time, they are not solving the problem they claim to solve. They never resolve this.\n\nWhat's actually new here is not much. The method is a direct application of CLIP-style contrastive cross-modal alignment plus Grad-CAM, and the authors cite WinCLIP and AnomalyGPT, which already apply vision-language models to IAD. There is no new architecture, no new loss, and no new theoretical insight. The paper is organized and readable, and the direction — leveraging textual descriptions for anomaly localization — is sensible. But the machinery is standard, and the experiments do not back the claims.\n\nThe soft spots are severe and mechanical. Eq. 7 uses inverse encoders f_v^{-1} and f_t^{-1} that are never defined, and there is no indication how reconstruction is computed. Baseline numbers in Table 1 do not match published values; for instance, WinCLIP's MVTec image AUC is typically around 91.8, not 93.1. The human evaluation protocol is unspecified — how many experts, what instructions, how the scale was used. The IoU results in Table 4 have no per-class breakdown and no statistical testing. The ablation study is only on MVTec, has no error bars, and the \"simple CNN\" row is an uninterpretable confound. There is no code, no training details, no hyperparameter settings.\n\nIn short, this paper is not ready for serious review. It might serve as a rough student project template, but the internal contradiction alone is enough to desk reject. I would not send it to referees; the flaws are load-bearing, not fixable with minor revision. Recommendation: reject without further consideration.","headline":"The paper's central claim is internally falsified by its own Table 1 on VisA pixel AUC, and the contrastive loss is incoherent under the stated normal-only setting; this is a desk reject.","tokens_in":10741,"tokens_out":2394,"would_cite":false,"duration_ms":22376,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"CLAD, a contrastive cross-modal training method, claims to outperform existing industrial anomaly detection methods on image-level and pixel-level benchmarks by aligning visual and textual embeddings.","keywords":["Large Vision-Language Models","Industrial Anomaly Detection","Contrastive Learning","Cross-Modal Alignment","Anomaly Localization","MVTec-AD","VisA","Interpretability"],"falsifier":"Train CLAD on the official MVTec-AD normal-only training split and compute the negative term in Eq. (3); if normal-instance embeddings spread apart under the loss while the paper's stated goal is to group them, then the reported 94.1 Image-AUC cannot be reproduced without adding anomalous pairs to training.","tokens_in":9686,"feed_emoji":"🔍","tokens_out":7242,"duration_ms":63036,"temperature":0.7,"pith_summary":"The paper introduces CLAD, a method that trains a vision encoder and a text encoder together so that an industrial product image is embedded close to its text description and normal instances are grouped together while anomalies are pushed apart. The claim is that this cross-modal contrastive alignment, followed by task-driven fine-tuning on industrial data, gives better anomaly detection and localization than existing unsupervised methods. On the MVTec-AD and VisA benchmarks, CLAD reports the highest Image-AUC and Pixel-AUC among the compared methods, and a human evaluation rates its localization heatmaps as more accurate than the baselines. The authors argue that this makes vision-language models a practical route to interpretable industrial quality control.","feed_headline":"Image-text alignment beats industrial anomaly baselines","feed_subtitle":"CLAD reports 94.1 Image-AUC on MVTec-AD, ahead of WinCLIP and PatchCore, with human-rated localization.","key_machinery":"The key mechanism is the contrastive cross-modal loss in Eq. (3), which pulls matched image-text pairs within a margin $\\alpha$ and pushes every non-matching pair beyond a margin $\\beta$, together with a reconstruction loss that preserves visual and textual details during fine-tuning. The visual encoder (ViT or CNN) and textual encoder (a pretrained transformer) map inputs to a shared embedding space, and the anomaly score $S(I,T)=\\exp(-\\|z_v-z_t\\|^2/\\sigma)$ turns the embedding distance into a similarity measure. For localization, the paper uses Grad-CAM to produce a heatmap of the image regions that most contribute to the mismatch.","core_discovery":"The central discovery is that aligning image and text embeddings with a contrastive loss is enough to separate normal from anomalous industrial images without relying on reconstruction error or memory banks. CLAD computes an anomaly score from the distance between the test image embedding and the text description embedding, so an image that does not match its expected textual prototype is flagged as anomalous, and Grad-CAM highlights the regions responsible. Across the 15 categories of MVTec-AD and 12 categories of VisA, the paper reports Image-AUC of 94.1 and 86.1 respectively, both above the strongest baseline WinCLIP, with Pixel-AUC of 95.3 and 96.2. The paper interprets these results as evidence that cross-modal alignment generalizes to unseen anomaly types and yields interpretable localizations.","pith_inferences":["The loss as written in Eq. (3) pushes all non-matching instances apart, so under the standard unsupervised setup with normal-only training images, the negative term would separate normal instances from each other; a natural follow-up experiment is to test whether the reported results require anomalous training pairs to be present.","A testable extension is to compare CLAD's text prototype score against a simpler zero-shot CLIP distance baseline with no fine-tuning, which would isolate the contribution of the contrastive training signal.","The same contrastive objective could be adapted to few-shot defect detection by adding a small set of textual descriptions of known defect types, turning the anomaly score into a multi-class defect classifier.","If the reconstruction loss term is removed, one can measure whether the remaining contrastive alignment alone preserves the reported localization quality, clarifying how much of the gain comes from reconstruction versus alignment."],"forward_implications":["If the reported numbers hold, a single contrastive alignment objective can outperform dedicated reconstruction and memory-bank pipelines for industrial defect detection and localization.","The text-guided anomaly score means new product types could be supported by writing a description of the normal object, without collecting defect samples for training.","The high human rating of the heatmaps suggests the same embeddings can serve as a visual explanation channel in factory quality-control interfaces.","Because the method scores by mismatch to the text prototype, it could in principle flag anomaly types never seen during training, which is the paper's stated goal."],"supporting_citations":[{"why":"Provides the primary benchmark, MVTec-AD, with 15 categories of industrial objects and defects used to evaluate detection and localization.","marker":"[3]"},{"why":"Provides the second benchmark, VisA, with 12 categories and 1,200 anomalous images used for evaluation.","marker":"[4]"},{"why":"Supplies the pretrained vision-language backbone that CLAD aligns and fine-tunes for industrial images.","marker":"[8]"},{"why":"PatchCore is a strong memory-bank baseline that CLAD compares against and must beat in Table 1.","marker":"[7]"},{"why":"WinCLIP is the closest competitor in Table 1, and its scores define the improvements CLAD claims.","marker":"[30]"},{"why":"AnomalyGPT is a prior LVLM-based industrial anomaly detection method that motivates CLAD's design.","marker":"[5]"},{"why":"PaDiM is a baseline whose Image-AUC and Pixel-AUC CLAD is compared against in the experiments.","marker":"[29]"}],"fun_headline_variants":["Image-text contrastive model beats industrial anomaly baselines","CLAD: aligning vision and language for robust anomaly detection","Contrastive cross-modal training excels at industrial anomaly spotting","LVLM-based CLAD sets new bar for anomaly detection and localization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training signal assumes the dataset provides both normal and anomalous instances to pull apart, but the benchmark is used in its standard unsupervised form with normal-only training images, so the paper never states where the negative pairs come from.","fun_headline_variants_meta":{"raw":{"variants":["Image-text contrastive model beats industrial anomaly baselines","CLAD: aligning vision and language for robust anomaly detection","Contrastive cross-modal training excels at industrial anomaly spotting","LVLM-based CLAD sets new bar for anomaly detection and localization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000553,"raw_usage":{"total_tokens":2599,"prompt_tokens":873,"completion_tokens":1726,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":489,"completion_tokens_details":{"reasoning_tokens":1658}},"tokens_in":489,"tokens_out":1726,"duration_ms":10210,"temperature":1.0,"reasoning_tokens":1658,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:53:35.845335+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train CLAD on the official MVTec-AD normal-only training split and compute the negative term in Eq. (3); if normal-instance embeddings spread apart under the loss while the paper's stated goal is to group them, then the reported 94.1 Image-AUC cannot be reproduced without adding anomalous pairs to training.","supporting_citations":[{"cited_title":"In: Proceedings of the IEEE/CVF conference on computer vision and pattern recogni tion","cited_arxiv_id":null,"evidence_quote":"Provides the primary benchmark, MVTec-AD, with 15 categories of industrial objects and defects used to evaluate detection and localization."},{"cited_title":"In: European Conference on Computer Vision","cited_arxiv_id":null,"evidence_quote":"Provides the second benchmark, VisA, with 12 categories and 1,200 anomalous images used for evaluation."},{"cited_title":"In: Meila, M., Zhang, T","cited_arxiv_id":null,"evidence_quote":"Supplies the pretrained vision-language backbone that CLAD aligns and fine-tunes for industrial images."},{"cited_title":"In: Proceedi ngs of the IEEE/CVF conference on computer vision and pattern recognition","cited_arxiv_id":null,"evidence_quote":"PatchCore is a strong memory-bank baseline that CLAD compares against and must beat in Table 1."},{"cited_title":"I n: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recogni tion","cited_arxiv_id":null,"evidence_quote":"WinCLIP is the closest competitor in Table 1, and its scores define the improvements CLAD claims."}],"review_version":1}