{"id":"8efb62a7-60f6-4b67-b256-508e79d60798","arxiv_id":"2507.17240","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A two-layer classifier trained on CONTRIQUE image-quality features achieves state-of-the-art accuracy on GenImage and DRCT-2M fake-image detection benchmarks.","lead":"This paper shows that internal features of image-quality-assessment models, originally built to score how natural an image looks, can also separate real photos from AI-generated ones. A small classifier on these features matches or beats previous state-of-the-art detectors on standard benchmarks.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DRCT-2M evaluation reuses the same MSCOCO real images across every generator subset, so CONTRIQUE's apparent SOTA on that dataset may reflect real-image memorization rather than cross-generator generalization.","rationale":"The paper's strongest claim is that a two-layer network on CONTRIQUE features achieves SOTA on GenImage and DRCT-2M. For that claim to hold, the benchmark numbers must be valid tests of cross-generator generalization. The DRCT-2M experiment is not such a test for the real class: §5.2 explicitly says the same MSCOCO real images are used across all generative subsets, and the classifier is trained on one of those subsets. Memorizing the training real images (or their features) would yield near-perfect real-class accuracy, directly inflating mAcc. All baselines share this protocol, but the relative gap (0.74 mAcc) is small and could flip if the methods exploit the overlap differently. The reader's weakest assumption about the Section 3.1 hypothesis is real but less load-bearing: even if the IQA-manifold explanation is wrong, the method could still work as a fingerprint detector. The validity of the DRCT-2M number, by contrast, is directly load-bearing for the SOTA claim. A held-out real-split test or per-class accuracy report settles it. We therefore keep the CONDITIONAL verdict; the paper should add this controlled evaluation, error bars, and code before the SOTA claim can be accepted.","tokens_in":14061,"tokens_out":7247,"duration_ms":73571,"concrete_test":"Re-evaluate on DRCT-2M using a real-image split that is disjoint from training: e.g., hold out 20% of the MSCOCO real images when constructing the SDv1.4 training set, and test on the same 20% held-out real images paired with each generator's fakes; then recompute mAcc. Additionally, report per-class real/fake accuracy and 95% bootstrapped confidence intervals over 5 training seeds. If CONTRIQUE's mAcc advantage over DRCT/UnivFD on DRCT-2M disappears (or falls below the GenImage gap of 0.56), the DRCT-2M SOTA claim is an artifact of real-image reuse. If the advantage persists, the confound is limited to raising absolute mAcc and the relative claim stands.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 5.2 states: 'the validation set contains the same set of real images from the MSCOCO dataset across all generative models.' Because the perceptual classifier is trained on the SDv1.4 subset of DRCT-2M, the real images seen during testing on every other subset (SDv2, SDXL, Turbo, LCM, ControlNet, DR variants) are identical to those used in training. mAcc is the average of real-class and fake-class accuracy, so a detector can inflate mAcc by memorizing the specific real images, independent of any understanding of generative artifacts. The same protocol was used for all compared baselines, but the relative ordering is still at risk: a feature space like CONTRIQUE's that is more sensitive to content/identity (rather than to generative fingerprints) would be preferentially rewarded. Thus the DRCT-2M SOTA result (91.23 vs 90.49 for DRCT/UnivFD) does not currently support the claim of generalization to unseen generative models. The GenImage result, where real images come from ImageNet splits, is less obviously confounded, but the margin over DRCT/UnivFD is only 0.56 mAcc and no error bars are provided. The central finding therefore rests on an insecure evaluation for at least one of the two headline datasets.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes \"perceptual classifiers\": frozen no-reference IQA backbones (CONTRIQUE, ReIQA, ARNIQA, HyperIQA, TReS) whose features are passed to a two-layer MLP trained with a combination of cross-entropy and margin-based contrastive loss on real, fake, real-reconstructed, and fake-reconstructed images. The classifiers are evaluated on GenImage, DRCT-2M, and UniversalFakeDetection, and the authors report state-of-the-art mean accuracy for CONTRIQUE (90.04 mAcc on GenImage, 91.23 mAcc on DRCT-2M), along with cross-dataset and distortion-robustness experiments.","tokens_in":14384,"tokens_out":7393,"duration_ms":69282,"significance":"If the evaluation is sound, the result is practically significant: CNN-based IQA backbones are lighter than CLIP-style transformers, already deployed in many image delivery pipelines, and their features can serve both quality prediction and fake-image detection. The paper offers broad comparisons across many generative models, cross-dataset transfer results, and fixed hyperparameters taken from prior work. The claim of state-of-the-art performance on GenImage and DRCT-2M is, however, weakened by a potentially confounded DRCT-2M protocol and by the absence of error bars; the abstract's robustness claim is also not supported by the paper's own Section 5.4. The work does not exhibit circularity: the classifier is trained on labeled real and fake data and evaluated on external test sets.","major_comments":[{"comment":"The DRCT-2M evaluation as described does not support the claim of generalization to unseen generators. Section 5.2 states that \"the validation set contains the same set of real images from the MSCOCO dataset across all generative models,\" and Section 4.1 trains the classifier on the SDv1.4 subset of DRCT-2M. Since mAcc is the average of real-class and fake-class accuracy, if those same MSCOCO real images are used during training, a detector can inflate its real-class accuracy on every test subset by memorizing those images, independently of generator-specific artifacts. The 0.74 mAcc margin over DRCT/UnivFD in Table 2 (91.23 vs 90.49) is then not strong evidence of cross-generator generalization. The same protocol applies to all baselines, but the relative ordering can still be biased if CONTRIQUE's feature space is more sensitive to content identity. Please state the explicit train/validation split, confirm whether real images overlap, and report per-class accuracies or re-evaluate on a disjoint real-image set.","section":"§5.2, Tables 2–3"},{"comment":"The abstract's claim of \"significant robustness against image degradations\" is contradicted by Section 5.4, which states that the CONTRIQUE classifier—the method achieving the headline SOTA numbers—\"exhibited higher vulnerability to image distortions when compared to ReIQA and DRCT/UnivFD classifiers.\" The Conclusion's statement that \"Our proposed models delivered better robustness against Gaussian blur\" is likewise not supported by Figure 6 or the surrounding text. Please report numeric mAcc values and standard deviations for each method at each degradation level, attribute robustness separately to each classifier, and revise the abstract and conclusion to match the actual results.","section":"§5.4, Fig. 6, Abstract, §6"},{"comment":"The headline SOTA margins are small (90.04 vs 89.48 mAcc on GenImage; 91.23 vs 90.49 on DRCT-2M) and are presented without error bars, repeated-seed variance, or significance tests. The per-subset numbers show that CONTRIQUE trails DRCT/UnivFD on BigGAN by 11.72 mAcc in Table 1 and on 7 of 16 DRCT-2M subsets in Table 2, so the aggregate advantage is driven by a subset of generators. Without variance estimates or a paired comparison, it is unclear whether the claimed advantage is stable. Please report at least three training runs with mean and standard deviation, or a paired test across subsets.","section":"Tables 1–3, §5.1"}],"minor_comments":[{"comment":"The metric \"mAcc\" is not precisely defined; please state whether it is the arithmetic mean of per-generator accuracies, the mean class accuracy, or overall accuracy, since the label affects how the tables are read.","section":"§4.2"},{"comment":"The sentence \"we used only data augmentation when studying specific degradation\" is unclear; please describe whether augmentations are applied at inference in the robustness experiments and how that interacts with the degradation being studied.","section":"§3.3"},{"comment":"The description of the DRCT-2M validation protocol should include the number of real images and an explicit statement of whether the same real images are reused across all generator subsets; this information is essential for interpreting Table 2.","section":"§5.2"},{"comment":"Reference [62] appears to be titled \"LSUN: Construction of a Large-scale Image Dataset...\", but the dataset described in the text is GenImage; please verify the citation and title.","section":"References"},{"comment":"There are small presentation errors: \"DRCT/CovnNext-B\" should be \"DRCT/ConvNext-B\", and the Table 2 column header \"LDMSDv1.4\" should be separated into \"LDM\" and \"SDv1.4\".","section":"Text and Tables"},{"comment":"For reproducibility, please report the hidden-layer width of the two-layer classifier and the feature dimension used for each IQA backbone.","section":"§3.3"}],"recommendation":"major_revision","confidential_remarks":"The core idea is attractive and the GenImage result may survive scrutiny, but the DRCT-2M real-image overlap issue should be checked carefully before publication; the robustness claim also needs to be reconciled with the paper's own Figure 6. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The useful result here is real: a two-layer head on frozen CONTRIQUE features gets 90.04 mAcc on GenImage, beating DRCT/UnivFD by 0.56 points, and it generalizes cross-dataset to UniversalFakeDetection better than DRCT/UnivFD when trained on GenImage (79.46 vs 76.77). That is a clean, lightweight contribution. The paper also does a broad comparison across five IQA backbones and three datasets, and the t-SNE in Figure 3 gives at least preliminary support for the idea that self-supervised IQA features separate real from fake images. This specific combination is not in the cited work, and it deserves credit.\n\nThe soft spot is real and the stress-test note lands. Section 5.2 says the DRCT-2M validation set uses the same MSCOCO real images across every generator subset. Since the classifier is trained on the SDv1.4 subset of DRCT-2M, those exact real images appear in training. mAcc is an average of real and fake accuracy, so a model can inflate it by memorizing real images rather than detecting generative artifacts. All baselines share the protocol, but that doesn't rescue the relative ranking: a feature space like CONTRIQUE's, which may be more content-sensitive, is preferentially rewarded. So the 91.23 vs 90.49 DRCT-2M result does not support generalization to unseen generators. The GenImage result, where real images are ImageNet splits, is less confounded and carries the paper.\n\nThe abstract's robustness claim is also overbroad. Section 5.4 says the CONTRIQUE classifier was more vulnerable to blur and JPEG than ReIQA and DRCT/UnivFD. That's the opposite of \"significant robustness\" as a blanket statement. No error bars, no significance tests, and no code are provided; the margins are small enough that this matters.\n\nNone of this is fatal. The central idea is plausible, the GenImage numbers are useful, and the evaluation flaw is fixable by re-testing on disjoint real images or at least dropping the DRCT-2M claim. A serious referee should see this, not a desk reject. I'd want the authors to correct the abstract, re-run DRCT-2M with non-overlapping real IDs, and add error bars before publication. The paper is worth a reading group discussion, mostly for the protocol lesson.","headline":"Frozen CONTRIQUE features with a small head genuinely beat DRCT/UnivFD on GenImage, but the DRCT-2M headline is compromised by shared real images and the abstract overstates robustness.","tokens_in":14892,"tokens_out":2441,"would_cite":true,"duration_ms":26242,"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":"The paper claims that a two-layer classifier trained on frozen features of no-reference image quality assessment models, especially CONTRIQUE, outperforms dedicated generative-image detectors on unseen generative models, reaching 90.04…","keywords":["image quality assessment","AI-generated image detection","deepfake detection","no-reference IQA","contrastive learning","generalization across generative models","diffusion models","perceptual features"],"falsifier":"A direct test would ablate the IQA-specific training: freeze a randomly initialized or ImageNet-pretrained ResNet-50 and train the same two-layer classifier on the same real/fake/reconstructed samples; if the GenImage and DRCT-2M accuracy advantage over the IQA backbone mostly disappears, then the paper's mechanism, and not the classifier architecture or reconstruction data, is what carries the result.","tokens_in":13924,"feed_emoji":"🖼️","tokens_out":6640,"duration_ms":63799,"temperature":0.7,"pith_summary":"Image quality assessment (IQA) models are built to measure how far an image has drifted from natural, undistorted appearance. This paper tries to show that those same models, used only as frozen feature extractors, already separate real photographs from AI-generated images well enough to build a state-of-the-art detector. The authors train a two-layer network on features from several no-reference IQA backbones and report the best results with CONTRIQUE, a self-supervised model, on the GenImage and DRCT-2M benchmarks. They also test robustness to Gaussian blur and JPEG compression and find the perceptual classifiers hold up well. If the claim holds, platforms already running IQA for streaming and storage can add fake-image detection on the same computation.","feed_headline":"Frozen quality-scoring features beat dedicated deepfake detectors","feed_subtitle":"A two-layer classifier on CONTRIQUE features tops GenImage and DRCT-2M while staying robust to blur and JPEG.","key_machinery":"The load-bearing object is the frozen feature space of no-reference IQA models, with CONTRIQUE's contrastively trained ResNet-50 backbone as the primary example. The classifier itself is a two-layer neural network trained with a margin-based contrastive loss plus a cross-entropy loss on real, fake, and Stable-Diffusion-reconstructed versions of both classes. Because the IQA backbone is frozen, the same features that feed a quality predictor also feed the real/fake detector, making the two tasks share a single forward pass.","core_discovery":"On the paper's own terms, the central discovery is that the feature space of IQA models—particularly the self-supervised ResNet-50 backbone of CONTRIQUE—is a strong and efficient space for distinguishing real from AI-generated images. A two-layer classifier frozen on those features achieves 90.04% mean accuracy on GenImage, 91.23% on DRCT-2M, and the best cross-dataset transfer among the compared methods, including when trained on one dataset and tested on another. The paper attributes this to the hypothesis that IQA models implicitly model the distribution of real images from a distortion perspective, so generated images appear as an out-of-distribution deviation. The evidence for the mechanism is a t-SNE plot showing clearer separation for CONTRIQUE and ReIQA features than for CLIP or supervised IQA models.","pith_inferences":["Because the frozen features are shared with quality prediction, an attacker who degrades an image to fool the detector would also lower the perceived quality score, coupling the two tasks in a way the paper does not explore.","A natural next experiment the paper does not run is fine-tuning the IQA backbone on real/fake pairs; that would likely improve accuracy further but would sacrifice the multi-task reuse that makes the approach cheap.","The t-SNE evidence is limited to one Stable Diffusion subset; a broader map across GAN and autoregressive models would test whether the separation is a general property of IQA features or specific to diffusion artifacts."],"forward_implications":["Platforms that already run no-reference IQA for streaming, storage, or quality-of-experience monitoring can add AI-generated-image detection with only a small trained classifier on top of the same frozen features.","Self-supervised IQA backbones generalize to unseen generative models better than supervised ones, so quality models should be chosen with pretraining strategy in mind when used for detection.","The GenImage-trained CONTRIQUE classifier transfers across datasets better than the compared state-of-the-art detectors, suggesting the perceptual features are not overfit to one generator family.","Robustness to Gaussian blur is competitive with prior art, while JPEG compression remains a weaker point, so deployment should consider compression-aware variants."],"supporting_citations":[{"why":"Supplies the CONTRIQUE backbone whose frozen features carry the main detector.","marker":"[27]"},{"why":"Supplies the ReIQA backbone used as the competitive second classifier.","marker":"[44]"},{"why":"Provides the reconstruction-based training setup, loss combination, and the DRCT-2M benchmark.","marker":"[8]"},{"why":"Defines the UnivFD baseline and the UniversalFakeDetection dataset the paper compares against.","marker":"[37]"},{"why":"Introduces the diffusion-reconstruction idea that motivates including reconstructed real and fake samples.","marker":"[53]"},{"why":"Provides the GenImage dataset used for training and evaluating generalization across eight generative models.","marker":"[62]"},{"why":"Offers the CLIP encoder whose feature space the paper contrasts with IQA features in Figure 3 and baselines.","marker":"[40]"}],"fun_headline_variants":["Perceptual features from IQA models crack GenAI detection","Two-layer classifier on IQA features beats dedicated detectors","CONTRIQUE features lead GenAI detection, robust to blur and JPEG","IQA features make a simple net state-of-the-art at fake image spotting","Quality metrics double as powerful AI-image detectors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The premise that IQA models trained on real and distorted images implicitly model the distribution of real images, so their feature space separates real from AI-generated content, is supported only by a t-SNE visualization on one dataset subset; if that separation does not generalize, the detector's gains would be a dataset-specific artifact.","fun_headline_variants_meta":{"raw":{"variants":["Perceptual features from IQA models crack GenAI detection","Two-layer classifier on IQA features beats dedicated detectors","CONTRIQUE features lead GenAI detection, robust to blur and JPEG","IQA features make a simple net state-of-the-art at fake image spotting","Quality metrics double as powerful AI-image detectors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000218,"raw_usage":{"total_tokens":1415,"prompt_tokens":896,"completion_tokens":519,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":435}},"tokens_in":512,"tokens_out":519,"duration_ms":5444,"temperature":1.0,"reasoning_tokens":435,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T14:53:23.990459+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A direct test would ablate the IQA-specific training: freeze a randomly initialized or ImageNet-pretrained ResNet-50 and train the same two-layer classifier on the same real/fake/reconstructed samples; if the GenImage and DRCT-2M accuracy advantage over the IQA backbone mostly disappears, then the paper's mechanism, and not the classifier architecture or reconstruction data, is what carries the result.","supporting_citations":[{"cited_title":"Image Quality Assessment using Contrastive Learning.IEEE Transactions on Image Processing, 31:4149–4161, 2022","cited_arxiv_id":null,"evidence_quote":"Supplies the CONTRIQUE backbone whose frozen features carry the main detector."},{"cited_title":"Re-IQA: Unsupervised Learning for Image Quality Assessment in the Wild.IEEE/CVF Conference on Computer Vision and Pat- tern Recognition, pages 5846–5855, 2023","cited_arxiv_id":null,"evidence_quote":"Supplies the ReIQA backbone used as the competitive second classifier."},{"cited_title":"DRCT: Diffusion Reconstruction Contrastive Training to- wards Universal Detection of Diffusion Generated Images","cited_arxiv_id":null,"evidence_quote":"Provides the reconstruction-based training setup, loss combination, and the DRCT-2M benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the UnivFD baseline and the UniversalFakeDetection dataset the paper compares against."},{"cited_title":"LSUN: Construction of a Large-scale Image Dataset using Deep Learning with Humans in the Loop.Ad- vances in Neural Information Processing Systems, 36, 2024","cited_arxiv_id":null,"evidence_quote":"Provides the GenImage dataset used for training and evaluating generalization across eight generative models."},{"cited_title":"Learning Transferable Visual Models From Natural Language Supervision.International Conference on Machine Learning, 139:8748–8763, 2021","cited_arxiv_id":null,"evidence_quote":"Offers the CLIP encoder whose feature space the paper contrasts with IQA features in Figure 3 and baselines."}],"review_version":1}