{"id":"5f259baa-d689-4725-89aa-c4ac9a2160db","arxiv_id":"2607.03862","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Semantics-irrelevant local-global texture relations, extracted after content suppression, serve as cross-paradigm forensic cues that let DTS-Det reach 99.6% accuracy on a new mixed-generation benchmark.","lead":"Existing AI-image detectors fail when moving between pure generation and image-conditioned editing; this paper builds a benchmark showing the gap and a detector that models texture relations after content suppression. Platforms that must catch both fully synthetic and edited fakes may gain a more stable forensic signal.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The residual extraction may not isolate truly semantics-irrelevant, model-related texture relations; residual content leakage or architecture-specific artifacts could drive the reported gains.","rationale":"The reader correctly isolates the PRNU-style residual pipeline (Sec. IV-A/B, Figs. 3–5) as the weakest assumption underwriting the strongest claim. The empirical package is otherwise strong: a new mixed-paradigm benchmark, clear failure diagnosis of prior cue families, multi-axis generalization (cross-paradigm, cross-dataset, cross-media, reconstruction and black-box attacks), and ablations that credit the relation-guided stream. No mathematical inconsistency or circular derivation appears. The residual-isolation step remains unproven for truly unseen generators and for residual content leakage; until that is stress-tested (or code/data released for independent verification), the claim that texture relations are the generalizable cross-paradigm cue stays conditional. My concern is essentially identical to the reader’s, so the verdict stays CONDITIONAL with no adjustment needed.","tokens_in":23065,"tokens_out":662,"duration_ms":6117,"concrete_test":"On a held-out set of 2K real + 2K generated images from two generators never seen in ConImageGen training or residual construction (e.g., Flux and a recent image-conditioned editor outside X2Edit/PicoBanana), recompute single-image traces T_i with Eqs. (3)–(6), form model-level patterns via Eq. (7), and measure PCE AUC of same-generator vs. different-generator matching. Simultaneously ablate DTS-Det by replacing R_tex(x) with either (a) raw RGB or (b) a simple high-pass residual; if PCE AUC falls near chance or the ACC gap vs. the SigLIP2+DINOv3 baseline collapses below 3 points, residual leakage/architecture specificity undermines the generalizable-cue claim.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on Sec. IV-A/B and Eqs. (2)–(7): multi-scale wavelet residual extraction + multi-window Wiener shrinkage + MLE aggregation (inspired by PRNU) is asserted to suppress semantic content and yield model-related texture patterns whose local-global relations (visualized in Fig. 3, PCE-validated in Fig. 5) constitute generalizable forensic evidence across paradigms. DTS-Det then encodes these traces via SegFormer into a token-level bias m (Eq. 13) that modulates DINOv3 attention (Eq. 14). If residual content leakage remains (high-frequency edges, lighting, or editing-induced structures that survive averaging) or if the patterns are merely architecture-specific fingerprints of the 13 ConImageGen generators rather than a paradigm-invariant cue, then the 10.5-point in-domain lift, 93+% cross-dataset numbers, and attack robustness are explained by stronger residual features + dual-backbone fusion rather than by the claimed transferable texture-relation evidence. The paper’s own PCE results already show architecture clustering (SD variants, Wukong), which is consistent with this alternative.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper studies cross-paradigm AI-generated image detection spanning image-free and image-conditioned generation. It introduces ConImageGen (364K images from 13 generators), shows that frequency, semantic, and prior semantics-irrelevant detectors transfer poorly across paradigms, and argues that PRNU-style residual aggregation reveals model-related texture patterns whose local-global relations form generalizable forensic evidence. DTS-Det encodes residual traces with SegFormer into a token-level bias that modulates DINOv3 attention (Eqs. 13–14), fused with a SigLIP2 semantic stream under LoRA adaptation. Empirically it reports 99.6% ACC on ConImageGen (+10.5 over the best baseline), 93.2%/94.1% on PicoBanana/RAID, 87.3% frame-level cross-media accuracy on GenVidBench, and strong robustness under JPEG/resizing, reconstruction, and transferable black-box attacks.","tokens_in":23426,"tokens_out":1405,"duration_ms":31240,"significance":"The cross-paradigm problem is well motivated and previously under-addressed; ConImageGen is a concrete, reusable contribution. The evaluation suite is unusually broad (in-domain, both transfer directions, external datasets, video frames, degradations, reconstruction, and black-box attacks), and gains over a strong dual-backbone fusion baseline (SigLIP2+DINOv3) are large and consistent across Tables I–IV and Fig. 7. If the residual-relation interpretation is substantiated, the work supplies a useful forensic perspective beyond frequency and semantic cues. Even with a more modest mechanistic claim, the benchmark and detector would remain valuable to media forensics.","major_comments":[{"comment":"Fig. 8 / Sec. V–VI ablation does not isolate the claimed mechanism. Variants A–C are SigLIP2, DINOv3, and their fusion; D is full DTS-Det. There is no control that (i) feeds residual-encoder features as ordinary concatenated features without the relation-guided attention bias of Eq. (14), or (ii) applies relation-guided attention without the residual operator R_tex of Sec. IV-B. Without these, the 10.5-point lift over SigLIP2+DINOv3 cannot be attributed to “texture-relation modeling” rather than simply adding a residual feature stream to a dual backbone. This is load-bearing for the paper’s central framing.","section":"Fig. 8, Sec. V-B, Eqs. (13)–(14)"},{"comment":"Sec. IV-A/B and Eqs. (2)–(7) assert that multi-scale wavelet residual extraction, multi-window Wiener shrinkage, and MLE aggregation isolate semantics-irrelevant, model-related texture patterns. Support is mainly visual (Fig. 3) and PCE matching (Fig. 5). PCE already shows architecture clustering (SD variants, Wukong), which is equally consistent with architecture-specific fingerprints. Please add quantitative checks that residuals are weakly tied to content (e.g., correlation of residual maps with CLIP/SigLIP features before vs after extraction; same-content different-generator residual similarity) and discuss residual content leakage (high-frequency edges, lighting, editing structure). If leakage remains, the “semantics-irrelevant texture relations” interpretation is overstated even if detection accuracy is high.","section":"Sec. IV-A/B, Eqs. (2)–(7), Figs. 3 and 5"},{"comment":"Tables II–IV report strong cross-paradigm, cross-dataset, and cross-media transfer, but the paper does not establish that held-out generators (PicoBanana, RAID-clean, GenVidBench) are architecturally novel relative to ConImageGen training models. Given the architecture clustering in Fig. 5, shared families could inflate apparent paradigm-invariant transfer. A short architectural-novelty analysis of the external sets (or an explicit leave-architecture-out split) is needed to support the claim that texture relations, rather than residual fingerprints of related generators, drive generalization.","section":"Tables II–IV, Fig. 5"}],"minor_comments":[{"comment":"Reference numbering is inconsistent: Sec. III-A cites GLIDE as [39], but the bibliography entry [39] is GenVidBench; GLIDE is [14]. Please renumber and recheck all generator citations.","section":"Sec. III-A, References"},{"comment":"Fig. 3 caption states “14 models” while ConImageGen is described as 13 generators; clarify whether a real-image aggregate or an extra model is included.","section":"Fig. 3"},{"comment":"Residual-extraction free parameters (wavelet levels J, window set Ω, noise std σ) and relation-guidance settings (λ, L_g=2) are only partially specified and not ablated; list defaults and sensitivity in the main text or appendix.","section":"Sec. IV-B, Sec. V-B, Sec. VI-A"},{"comment":"Tables I–IV report point estimates only; add multi-seed means/std or bootstrap intervals for the main averages, especially the 10.5-point claim.","section":"Tables I–IV"},{"comment":"Eq. (7) reuses intensity-weighted MLE from PRNU (T_i ⊙ X_i). Briefly justify why intensity weighting is appropriate for synthetic texture traces rather than unweighted residual averaging.","section":"Eq. (7)"},{"comment":"The abstract’s “for the first time” claim for semantics-irrelevant texture visualization should be tempered relative to prior PRNU/Noiseprint-style residual work, or the novelty should be stated more precisely (cross-paradigm generative models).","section":"Abstract, Sec. IV-A"}],"recommendation":"major_revision","confidential_remarks":"Empirically this is a strong package; with the requested mechanism ablations and residual-validation checks it is a plausible accept. The main risk is overclaiming “texture relations as a new forensic principle” if residual features alone explain most of the lift—encourage a more modest framing if that is what the ablations show. Scope fits a solid CV/forensics venue."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is straightforward. Cross-paradigm detection (image-free vs image-conditioned) is a real gap, ConImageGen fills it cleanly, and DTS-Det posts large, consistent gains over strong baselines including fused DINOv3+SigLIP2. That is the part worth your time.\n\nWhat is new is not another residual detector. It is the explicit diagnosis that existing frequency, semantic, and low-level cues fail when you move between paradigms, the first multi-model spatial visualizations of aggregated texture maps after content suppression (Fig. 3), the PCE architecture-correlation check (Fig. 5), and the architectural move from matching explicit traces to injecting a learned local-global texture-relation bias into DINOv3 attention. Tables I–IV and the robustness suite (degradation, reconstruction, black-box) show the lift is not confined to one split: 99.6 % in-domain, 93–94 % cross-dataset, 87 % zero-shot video frames, and still high under reconstruction and transfer attacks. Ablations isolate the relation-guidance term. Citations are appropriate; the dual-backbone fusion baseline already controls for “just use better pretrained features.”\n\nThe soft spot is exactly the one the stress-test flags, and it is real but not fatal. The PRNU-style wavelet + multi-window Wiener + MLE pipeline is asserted to isolate semantics-irrelevant, model-related texture whose relations transfer. PCE already shows architecture clustering (SD variants, Wukong), so some of the signal is fingerprint-like. Residual content leakage or generator-family artifacts could still be doing part of the work. That does not erase the empirical gains or the cross-media result, but it means the interpretive claim “texture relations as paradigm-invariant forensic evidence” is stronger than the isolation proof. No public code or exact splits yet, and no error bars. Free parameters (LoRA, λ, Lg=2, wavelet settings) are ordinary for this literature.\n\nThis is for people building or evaluating practical AIGI detectors who care about image-conditioned editors and robustness. It is not a theory paper. I would send it to peer review; the problem, data, and multi-axis numbers justify referee time even if the residual story needs tightening. Engage if you work on detection; skim the figures and tables if you only need the benchmark.","headline":"Solid empirical forensics paper: new mixed-paradigm benchmark + residual-relation detector that actually moves the numbers; the PRNU-style isolation claim is the soft spot, not a collapse.","tokens_in":24020,"tokens_out":578,"would_cite":true,"duration_ms":5794,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Semantics-irrelevant texture relations, not explicit artifacts, give detectors a cue that works across image-free and image-conditioned generation.","keywords":["AI-generated image detection","cross-paradigm generalization","texture relations","semantics-irrelevant cues","image-conditioned generation","forensic residual analysis","DTS-Det","ConImageGen"],"falsifier":"Train DTS-Det only on ConImageGen, then measure accuracy on a large set of images from a brand-new architecture (neither image-free nor image-conditioned family used in the paper); if accuracy collapses toward chance while residual maps no longer show the claimed local-global structure, the central cue fails.","tokens_in":23992,"feed_emoji":"🔍","tokens_out":759,"duration_ms":9222,"temperature":0.7,"pith_summary":"Most AI-image detectors were built for images made from noise or text alone. Image-conditioned generators that edit or guide from a real photo are now common, and detectors trained on one style fail on the other. The paper builds ConImageGen, a 364K-image benchmark spanning both paradigms, and shows existing frequency, semantic, and low-level cues do not transfer. By averaging many images to suppress content, the authors expose model-linked texture patterns that still differ from real photos and that share structured local-to-global relations. They therefore stop hunting for fixed artifacts and instead train DTS-Det to encode those relations (plus light semantic cues). The detector reaches 99.6% accuracy on the benchmark, stays above 93% on external datasets, and holds up under reconstruction and black-box attacks, arguing that texture relations are a more stable forensic signal than the cues used today.","feed_headline":"Texture relations beat artifacts for spotting AI images","feed_subtitle":"A residual cue works across text-only and photo-guided generators, holding 93%+ on unseen data.","key_machinery":"DTS-Det: a dual-stream network whose texture-relation stream extracts wavelet residuals, encodes local-global texture relations with a hierarchical encoder, and injects them as continuous bias into relation-guided attention, fused with a lightweight semantic stream.","core_discovery":"After content suppression, generative models from both image-free and image-conditioned paradigms leave visible, model-related texture patterns whose structured local-global relations differ from real images; modeling those relations, rather than memorizing explicit artifacts, yields a generalizable cross-paradigm detector.","pith_inferences":["If texture relations track synthesis architecture more than training data, open-weight model releases could be fingerprinted by their aggregated residual maps alone.","The same content-suppression step might expose whether watermarking or post-processing erases local-global structure or only high-frequency noise.","A minimal detector that uses only the texture-relation stream (no semantic backbone) would test how much of the claimed gain is truly semantics-irrelevant."],"forward_implications":["Cross-paradigm detection becomes a first-class evaluation axis rather than an afterthought.","Detectors can shift from hunting generator-specific fingerprints to learning relation statistics that survive re-generation.","Reconstruction and transfer attacks that erase fixed artifacts leave relation structure partly intact, raising the bar for practical evasion.","The same residual-plus-relation pipeline can be tested as a frame-level signal for synthetic video without video-specific training.","New generators should be checked for whether their texture-relation signature still separates from real images before deployment."],"fun_headline_variants":["Local-global texture relations detect AI images across paradigms","Content-suppressed textures expose AI fakes from both generators","Modeling texture links beats artifacts in cross-paradigm detection","Structured texture patterns generalize AI image spotting","Ghost textures under content reveal cross-paradigm AI cues"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The multi-scale residual extraction and aggregation truly isolate stable, semantics-free texture relations that stay diagnostic for generators and paradigms never seen in training.","fun_headline_variants_meta":{"raw":{"variants":["Local-global texture relations detect AI images across paradigms","Content-suppressed textures expose AI fakes from both generators","Modeling texture links beats artifacts in cross-paradigm detection","Structured texture patterns generalize AI image spotting","Ghost textures under content reveal cross-paradigm AI cues"]},"model":"grok-4.5","effort":"low","cost_usd":0.005134,"raw_usage":{"total_tokens":1497,"prompt_tokens":866,"num_sources_used":0,"completion_tokens":79,"cost_in_usd_ticks":51340000,"prompt_tokens_details":{"text_tokens":866,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":552,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":866,"tokens_out":79,"duration_ms":4531,"temperature":1.0,"reasoning_tokens":552,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-11T23:25:31.540471+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train DTS-Det only on ConImageGen, then measure accuracy on a large set of images from a brand-new architecture (neither image-free nor image-conditioned family used in the paper); if accuracy collapses toward chance while residual maps no longer show the claimed local-global structure, the central cue fails.","supporting_citations":[],"review_version":1}