{"id":"5df9d46b-e0bb-4080-b690-f4928764363a","arxiv_id":"2501.13475","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"LDR-Net detects AI-generated images by combining a gradient residual with a random-weight local pattern descriptor, reporting top accuracy across GAN and diffusion generators.","lead":"The paper presents LDR-Net, a detector that flags AI-generated images by analyzing local smoothness and pixel-pattern anomalies. It reports the best or near-best accuracy on benchmarks covering GANs and diffusion models, although the gains over prior methods are modest.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cross-generator SOTA claim lacks a real-only control: the reported ACC/AP could reflect real-domain separation (LSUN vs. ImageNet/LAION) rather than detection of generator artifacts.","rationale":"The reader's weakest assumption names the same issue: the generalization results depend on the universality of local smoothing/pattern artifacts, but no controlled test separates artifact universality from dataset statistics. My stress-test sharpens this into a falsifiable condition: the classifier must not distinguish real images across domains. The absence of real-only accuracy is particularly important because pooled AP can remain high if fake images are ranked confidently even when real images are also misclassified. The proposed concrete test settles the concern directly. I do not move the verdict because the reader already set CONDITIONAL, and this concern reinforces rather than replaces that conditional status. If the real-vs-real control fails, the verdict should move to REJECT; if it passes, the cross-generator claim is substantially strengthened.","tokens_in":15481,"tokens_out":7008,"duration_ms":69814,"concrete_test":"Run the trained LDR-Net on the real-only partitions of DiffusionForensics, Ojha, and Self-Synthesis, and report real-class ACC/AP per source; then run a real-vs-real control: classify real images from the LSUN training categories against real images from ImageNet, LAION, and held-out LSUN categories at matched resolution and compression. If real-only ACC is materially below 50% on any test source, or if the real-vs-real separation is substantially above chance, the cross-generator claim is confounded by domain statistics rather than generator artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is generalization across unseen generators, but the evaluation never isolates the real class across domains. LDR-Net is trained on ProGAN images from four LSUN categories (cars, cats, chairs, horses; Sec. 4.1), while the test sets pair generated images with real images from other LSUN splits, ImageNet, and LAION. LGA (Sec. 3.1) is a high-frequency residual after Gaussian smoothing, and LVP (Sec. 3.2) is an LBP-style pattern code; both are known to be sensitive to sharpness, resolution, noise, and compression. If the network is separating the real training distribution from each test distribution, rather than separating real from generated, the pooled ACC and AP in Tables 2-5 could remain high even while per-source true-negative accuracy collapses. The paper reports no real-only accuracy per test source and no control where test real images are replaced by LSUN training-category real images. This is precisely the artifact-universality vs. dataset-statistics confound: a necessary condition for the universal-footprint claim is that real images from any source are classified as real at chance-level or better.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LDR-Net, a detection framework for AI-generated images that combines two handcrafted local features: LGA, a high-frequency residual obtained by subtracting a Gaussian-smoothed gradient magnitude map from the original gradient magnitude map, and LVP, a scalar encoding of thresholded local intensity differences with randomly chosen fixed weights. These feature maps are concatenated and fed into a pruned ResNet classifier. The model is trained on ProGAN-generated images and LSUN real images from four categories (cars, cats, chairs, horses) and evaluated on four benchmarks: ForenSynths, DiffusionForensics, Ojha, and Self-Synthesis. The paper reports average ACC/AP values that match or exceed recent baselines such as NPR and MI Net, together with ablations, a sensitivity study of the Gaussian smoothing parameter sigma, robustness to Gaussian blur and resizing, and CAM visualizations.","tokens_in":15698,"tokens_out":5258,"duration_ms":48134,"significance":"If the reported cross-generator accuracy holds, the paper would demonstrate that simple local discrepancy features—gradient residual statistics and local binary pattern complexity—can generalize across unseen GANs and diffusion models at a level competitive with recent learned methods. The empirical scope is substantial: four benchmarks, eight GAN sources, eight diffusion sources, and self-synthesized images, with comparisons to a current set of baselines. The training protocol follows the now-standard ProGAN/LSUN setup, which makes the results directly comparable to prior work. The main weakness is that the universal-footprint interpretation is not yet isolated from real-domain shift: because the real test images come from different datasets than the training real images, the reported ACC/AP could partly reflect domain separation rather than artifact detection. The paper therefore needs a real-only control before the central generalization claim can be accepted.","major_comments":[{"comment":"The cross-generator generalization claim is not supported by a real-only control. The classifier is trained on real LSUN images from four categories and ProGAN fakes, while the test sets pair fake images with real images drawn from other LSUN splits, ImageNet, LAION, and social-platform data. If the model is separating the real training domain from the test real domains rather than real from generated, the pooled ACC/AP in Tables 2–5 could remain high even if true-negative accuracy on ImageNet/LAION real images is near zero. Because only per-source ACC/AP is reported and no per-source true-negative rates or confusion matrices are given, the reader cannot distinguish the artifact-universality hypothesis from a dataset-statistics confound. Please report per-source real-class accuracy and include a control in which the real test images are drawn from the same LSUN distribution as the training set; this is a necessary condition for the claimed universal footprint.","section":"§4.1, Tables 2–5"},{"comment":"The Gaussian smoothing parameter sigma is selected by evaluating on the ForenSynths dataset, which is also one of the test sets used for the headline cross-GAN results in Table 2. This makes the reported ForenSynths numbers optimistically biased and complicates comparison with baselines whose hyperparameters were not tuned on the same test set. Please move sigma selection to a validation split, or report validation-based selection, and state explicitly whether the same sigma is used for all other test sets without any test-set feedback.","section":"§4.5, Table 6"}],"minor_comments":[{"comment":"The unqualified claim of state-of-the-art performance is overstated: in Table 2 LDR-Net ties with NPR on mean ACC (90.8%) and is below Ojha on mean AP (95.3% vs. 98.3%). Please qualify the claim as matching or surpassing on ACC while being second-best on AP.","section":"Abstract and §4.3"},{"comment":"Equation (9) uses N both as the upper summation limit and inside 'N = 7', which is confusing; define the eight weights W_0,...,W_7 explicitly. Also, the random selection of these weights is not specified with a distribution or seed, which hurts reproducibility.","section":"Eq. (9)"},{"comment":"The text says the method is tested on 'DIRE [Wang et al., 2023] and Ojha [Wang et al., 2023]', but Table 3 is headed 'DiffusionForensics' and the Ojha dataset reference should be [Ojha et al., 2023]. Please align the dataset names and references.","section":"§4.4"},{"comment":"In the F3Net row, the AP value '99,5' uses a comma as the decimal separator; please use '99.5' for consistency.","section":"Table 3"},{"comment":"The contribution text describes LVP as capturing 'low-frequency inconsistencies', but LVP is computed from per-pixel intensity differences and is not a low-frequency feature. Rephrase to avoid this inconsistency.","section":"§1 and §3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope for a computer vision venue and the experimental comparison is broad. The main issue is the real-domain confound: without per-source real accuracies or a same-domain real control, the central claim of a universal generator-artifact footprint is not yet established. The sigma-selection-on-test issue is also a standard but important methodological flaw that should be fixed. The novelty is moderate—two handcrafted local features feeding a standard classifier—but the empirical results are competitive and worth publishing if the controls are added. I do not see evidence of circularity or fabrication; the concern is missing evidence, not internal inconsistency."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a workmanlike empirical paper with a simple front end that roughly matches or slightly beats NPR on diffusion benchmarks, and a load-bearing confound the authors never address. The components are not new—LGA is a high-pass residual on Sobel magnitudes, LVP is an LBP variant with random weights—but the combination and the breadth of evaluation are new. The ablation shows both modules contribute, and the robustness results under blur and resizing are the strongest part of the paper.\n\nThe main problem is the real-image side. Training real images are LSUN (cars, cats, chairs, horses). Test real images in the diffusion benchmarks come from ImageNet and LAION. If the model is separating “LSUN-like” from “everything else,” the high ACC on unseen generators could be domain discrimination rather than artifact detection. The paper reports only pooled ACC/AP, no per-source true-negative accuracy, and no control where test real images are replaced by LSUN training-category real images. That is exactly the control needed to support the universal-footprint claim. The ForenSynths results are less exposed if its real set is also LSUN, but the paper never states the composition of its real test splits, and the cross-domain claim still needs the control. The stress-test note lands; this is a real confound, though likely fixable.\n\nOther soft spots, in proportion. Sigma is selected using the test set (Table 6), which inflates the headline numbers by a small margin. There are no error bars or multiple seeds. Baseline numbers are largely copied from NPR, so unfair comparisons can creep in. The random LVP weights and pruning details are unspecified, so exact reproduction is impossible without code. And the “state-of-the-art” phrasing oversells: on ForenSynths the ACC ties NPR at 90.8 and AP trails Ojha by three points. The honest claim is “competitive ACC on GANs, best ACC on diffusion.”\n\nThe math itself is elementary and correct; the derivations are straightforward. The citation pattern is normal for this area.\n\nI would send this to a serious referee. The method is simple enough to be reproducible, the benchmarks are broad, and the confound is addressable in revision. I would not desk-reject. The referee should ask for per-source real accuracy, a same-domain control, a validation-based sigma selection, and full specification of the random weights and seeds. If those come out clean, LDR-Net becomes a useful practical baseline. I would also consider citing the LGA/LVP combination once code is available.","headline":"A simple, cheap local-feature detector with broad benchmarks but a domain confound that needs a control before the cross-generator claim holds.","tokens_in":16231,"tokens_out":3599,"would_cite":true,"duration_ms":36264,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"One detector can catch GAN and diffusion images from their localized smoothing and pattern-uniformity artifacts.","keywords":["AI-generated image detection","local gradient autocorrelation","local variation pattern","cross-generator generalization","GAN-generated image detection","diffusion model detection","image forensics","localized discrepancy representation"],"falsifier":"Find a generator whose images have natural-looking local gradients and high directional-code entropy — for example, a diffusion model trained with an explicit texture-preserving loss — and test LDR-Net on it after training only on ProGAN images; if accuracy drops to chance while human observers still judge the images synthetic by other means, the local-discrepancy premise is falsified. A cheaper check is to compare LGA and LVP feature distributions on real versus generated images across many generator families and look for an unseen family whose feature distributions overlap real images.","tokens_in":15264,"feed_emoji":"🕵️","tokens_out":8845,"duration_ms":68652,"temperature":0.7,"pith_summary":"The paper tries to show that AI-generated images share a detectable local defect: generative models over-smooth small regions and produce unnaturally regular pixel patterns, whereas real images keep complex, random local texture. On that basis it proposes LDR-Net, a detector built from two complementary feature extractors — local gradient autocorrelation (LGA) for smoothing anomalies and local variation pattern (LVP) for pattern uniformity — whose outputs are concatenated and classified by a pruned residual network. Trained only on ProGAN images from four LSUN categories, the network reports average accuracies of 90.8% on ForenSynths, 96.0% on DiffusionForensics, 93.1% on Ojha, and 86.2% on Self-Synthesis, matching or exceeding published baselines on images from unseen GANs and diffusion models. If the claim holds, one compact detector trained once on a single generator family could serve as a practical first-line filter for synthetic images across evolving generative architectures.","feed_headline":"Too-smooth textures and flat pixel patterns give AI images away","feed_subtitle":"Trained only on ProGAN images, it still flags unseen GANs and diffusion models at 86–96 percent accuracy.","key_machinery":"The load-bearing mechanism is the paired representation of localized discrepancy. LGA builds a residual between the Sobel gradient magnitude and its Gaussian-smoothed version, isolating the high-frequency smoothing anomalies that generated images tend to erase. LVP turns each 3×3 neighborhood into eight binary direction codes, one per neighbor indicating whether the center pixel is brighter, weights each direction with a distinct real number, and sums them into a single value whose distribution reflects local pattern complexity; generated images produce narrower, more uniform distributions. The two feature maps are concatenated and fed to a pruned residual classifier, so the detector sees edge-texture anomalies and pixel-distribution regularity at the same time.","core_discovery":"The central discovery is that local discrepancy, not global statistics, separates generated from real images. Generated images carry excessive smoothness, blurred textures, and overly uniform small-scale pixel variation because generators apply smoothing to force visual coherence, while real images retain complex, diverse local patterns and natural randomness. LDR-Net makes this measurable in two ways: it computes the Sobel gradient magnitude, smooths it with a Gaussian kernel, and subtracts the smoothed version to leave high-frequency anomalies (LGA); and it encodes each pixel's intensity relation to its eight neighbors as a weighted directional code whose value distribution is narrower and simpler for generated images (LVP). Concatenating LGA and LVP features and training a pruned residual classifier on ProGAN images yields, the paper reports, a detector that outperforms or matches prior methods on unseen GANs and diffusion models and keeps high accuracy under Gaussian blur and resizing.","pith_inferences":["An implication the paper leaves implicit is that LGA and LVP are simple convolutional operations that could run before or beside a deep classifier, making a lightweight screening tool feasible on devices with limited compute.","The paper's logic predicts that a generator trained with an explicit penalty on local over-smoothing and pattern uniformity — for example, a loss that pushes gradient autocorrelation and directional-code entropy toward real-image values — should defeat this detector faster than a generator that only matches global statistics.","A boundary of the reported generalization is that training uses only four LSUN object categories; if local artifact statistics differ by content domain, a detector trained on objects may not transfer to faces or medical imagery, a split the paper does not test.","The robustness results suggest a testable extension: measuring how accuracy decays under JPEG compression, which the paper says is reported in supplementary material, would show whether the local features survive the most common real-world re-encoding."],"forward_implications":["A detector trained once on ProGAN images from four LSUN categories can transfer to GAN families such as StyleGAN, BigGAN, CycleGAN, StarGAN, GauGAN, and DeepFake without retraining, at the reported accuracy.","The same training transfers to diffusion families such as ADM, DDPM, IDDPM, LDM, PNDM, VQ-Diffusion, and Stable Diffusion, with average accuracy around 96% on the DiffusionForensics set.","Because the features are local and relational rather than absolute pixel values, the method survives Gaussian blur and resizing better than the compared baselines, keeping average accuracy near 92% and 95% respectively on DiffusionForensics.","Longer diffusion sampling and text-to-image generators do not erase the artifacts: on the Self-Synthesis set, LDR-Net stays above 86% average accuracy against diffusion models and Midjourney/DALL-E content.","The two modules are complementary: each alone reaches about 85–86% average accuracy on ForenSynths, while the combined network reaches 90.8%, indicating they catch distinct artifact dimensions."],"supporting_citations":[{"why":"Supplies the ForenSynths dataset, the ProGAN training images, and the CNNDetection baseline that LDR-Net is trained on and compared against.","marker":"[Wang et al., 2020]"},{"why":"Provides the NPR baseline, the Self-Synthesized test set with 1000-step diffusion and text-to-image samples, and the choice of four training categories.","marker":"[Tan et al., 2024]"},{"why":"Provides the LSUN real images paired with ProGAN-generated images for training.","marker":"[Yu et al., 2015]"},{"why":"Supplies the Ojha cross-diffusion test set and a universal-detector baseline whose average precision LDR-Net is compared against.","marker":"[Ojha et al., 2023]"},{"why":"Supplies the DIRE/DiffusionForensics dataset with eight diffusion models used for cross-diffusion evaluation and robustness tests.","marker":"[Wang et al., 2023]"},{"why":"Provides the residual-convolution backbone that the classifier prunes and trains to distinguish the concatenated LGA and LVP features.","marker":"[He et al., 2016]"}],"fun_headline_variants":["Local anomalies reveal AI images: LDR-Net spots them","Smoothness and texture quirks expose generated photos","Detect fakes by local pattern oddities, not global stats","LDR-Net: new detector catches AI images via local discrepancies","Small-scale pixel irregularities give away AI-generated images"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire cross-generator claim rests on the assumption that local over-smoothing and pattern uniformity are universal footprints of generative models, so that a classifier trained only on ProGAN images from four LSUN categories transfers to unseen GANs and diffusion models; the paper provides no control experiment that separates artifact universality from dataset statistics.","fun_headline_variants_meta":{"raw":{"variants":["Local anomalies reveal AI images: LDR-Net spots them","Smoothness and texture quirks expose generated photos","Detect fakes by local pattern oddities, not global stats","LDR-Net: new detector catches AI images via local discrepancies","Small-scale pixel irregularities give away AI-generated images"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3129,"prompt_tokens":947,"completion_tokens":2182,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":563,"completion_tokens_details":{"reasoning_tokens":2099}},"tokens_in":563,"tokens_out":2182,"duration_ms":13137,"temperature":1.0,"reasoning_tokens":2099,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:53:41.018377+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Find a generator whose images have natural-looking local gradients and high directional-code entropy — for example, a diffusion model trained with an explicit texture-preserving loss — and test LDR-Net on it after training only on ProGAN images; if accuracy drops to chance while human observers still judge the images synthetic by other means, the local-discrepancy premise is falsified. A cheaper check is to compare LGA and LVP feature distributions on real versus generated images across many generator families and look for an unseen family whose feature distributions overlap real images.","supporting_citations":[{"cited_title":"Cnn- generated images are surprisingly easy to spot","cited_arxiv_id":null,"evidence_quote":"Supplies the ForenSynths dataset, the ProGAN training images, and the CNNDetection baseline that LDR-Net is trained on and compared against."},{"cited_title":"Rethink- ing the up-sampling operations in cnn-based generative network for generalizable deepfake detection","cited_arxiv_id":null,"evidence_quote":"Provides the NPR baseline, the Self-Synthesized test set with 1000-step diffusion and text-to-image samples, and the choice of four training categories."},{"cited_title":"Dire for diffusion-generated image detec- tion","cited_arxiv_id":null,"evidence_quote":"Supplies the DIRE/DiffusionForensics dataset with eight diffusion models used for cross-diffusion evaluation and robustness tests."}],"review_version":1}