{"id":"80bd995e-4b09-4e4a-b233-2ccc99ac3dc8","arxiv_id":"2607.02886","paper_version":1,"verdict":"ACCEPT","confidence":"HIGH","novelty_score":6.5,"correctness_risk":"low","formal_verification":"none","parameter_count":3,"one_line_summary":"Training-free patch-token signals (TTR + LSMI) detect fully generated and partially edited videos at 0.1% FPR better than supervised and training-free baselines.","lead":"SPLIT is a training-free detector that flags AI-generated and partially edited videos by measuring patch-level temporal roughness and spatial motion incoherence in frozen vision-encoder features. It targets the ultra-low false-positive regime needed for real services and reports the highest Fake Recall at 0.1% FPR across three benchmarks.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified","rationale":"The paper's strongest claim is empirical and is backed by three complementary benchmarks, real-only calibration with cross-real transfer, thorough ablations (patch-level vs. global, TTR step size, gamma, encoder family), post-processing robustness, and public code. The assumption that the engineered patch-token cues generalize is the natural point of residual risk for any training-free detector, but the authors already confront it with out-of-distribution generators and multiple frozen backbones; the remaining risk is ordinary future-generator risk rather than an internal flaw. I therefore leave the reader's ACCEPT (high confidence) unchanged and treat the concrete test as a reproducibility sanity check rather than a falsification of a soft assumption.","tokens_in":44793,"tokens_out":452,"duration_ms":5525,"concrete_test":"Independently re-run the FakeParts Overall Fake Recall@0.1% pipeline (XCLIP-B/16, gamma=8, cross-real harmonic mean of ROVI/MSR-VTT thresholds) on the public code and data; if the reproduced overall figure falls outside the reported 95% bootstrap CI [72.42%, 76.14%], the headline numbers would need re-examination.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is well-supported by the reported evidence. The reader's weakest assumption (that TTR/LSMI remain generator-agnostic under fixed gamma and real-only calibration) is the right soft spot, but the paper already stress-tests it: ViF-Bench covers 2024–2025 generators, encoder ablations (Tab. 7) show the gain is not backbone-specific, gamma is fixed after a single held-out choice and saturates (Fig. 4 + App. E), and cross-real transfer (Tab. 4) keeps actual FPR near target. Inpainting is weak (2.48% at 0.1% FPR) yet still ranks second and is disclosed. No hidden circularity, derivation gap, or evaluation leak is apparent; the service-aligned protocol itself is a contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes SPLIT, a training-free detector for fully AI-generated and partially edited videos that scores frozen vision-encoder patch tokens with two complementary signals: Two-step Temporal Roughness (TTR; Eqs. 2–5), a log-ratio of accumulated one-step versus normalized two-step feature path lengths, and Local Spatial Motion Incoherence (LSMI; Eqs. 6–9), the mean magnitude of spatial gradients of a feature-space motion field. The signals are fused as TTR^γ · LSMI with a single fixed γ=8 (Eq. 10) and thresholded under a service-aligned protocol: Fake Recall at FPR ∈ {0.1%, 1%, 5%} with real-only calibration and cross-real threshold transfer. On FakeParts, GenVideo, and ViF-Bench, SPLIT reports the highest Fake Recall at FPR=0.1% (74.45%, 85.16%, 84.52% overall under cross-real transfer), outperforming supervised baselines and training-free D3, with supporting ablations (components, γ, encoders), post-processing robustness, bootstrap CIs, and AUROC tables.","tokens_in":45011,"tokens_out":1067,"duration_ms":10530,"significance":"If the results hold, the work is significant on two fronts. First, it shows that localized spatiotemporal incoherence on frozen patch tokens can detect both full synthesis and sparse partial edits without task-specific training, with large gains at the ultra-low-FPR regime that matters for deployment. Second, the evaluation protocol itself—Fake Recall at fixed FPR with real-only calibration and cross-real transfer—is a useful corrective to AUROC-centric reporting in this area. Strengths include public code, multi-benchmark generator- and category-stratified tables, encoder-agnostic gains (Tab. 7), fixed-γ saturation (Fig. 4, App. E), cross-real FPR stability (Tab. 4), and bootstrap CIs that separate SPLIT from D3. The main practical caveat is weak inpainting recall, which the authors disclose.","major_comments":[{"comment":"§4.2 and Tab. 1: Inpainting Fake Recall at FPR=0.1% is only 2.48% (second-best among compared methods). Because FakeParts is a central claim for partial-edit detection, the manuscript should more explicitly bound when patch-level TTR/LSMI is expected to fail (e.g., very small spatial support) and, if possible, quantify recall as a function of edited-area fraction or provide a simple spatial-max/top-k pooling variant so readers can see whether the failure is fundamental or aggregation-related.","section":null},{"comment":"§3.5 and §5.2: γ=8 is chosen once on a small held-out split and then frozen. The paper already shows saturation on FakeParts/GenVideo/ViF-Bench, but the held-out split is not named or sized. A one-sentence statement of which data were used for that single choice (and confirmation that no test-set FPR targets were used) would close the only remaining free-parameter concern for reproducibility.","section":null}],"minor_comments":[{"comment":"Fig. 2 fusion formula uses nonstandard notation (TTR! × LSMI, γ=8); align it with Eq. (10) for consistency.","section":null},{"comment":"Eq. (3): the normalization factor (T−1)/(T−2) is clear for T≥3, but a brief note on minimum clip length after 8 fps sampling would help implementers.","section":null},{"comment":"Tab. 3 header says FPR=0.1% only; a pointer in the caption to App. Tabs. 8–9 for 1%/5% would improve navigation.","section":null},{"comment":"§4.1 Implementation Details: crop 10% from the longer side is underspecified (center vs. random); state the exact policy used for all reported numbers.","section":null},{"comment":"Related Work: a short explicit contrast with ReStraV’s frame-level straightening vs. SPLIT’s patch-level 1-step/2-step ratio would sharpen positioning.","section":null},{"comment":"Appendix C.1: the multi-step slope fit for K>2 is interesting; a one-line intuition why longer steps dilute the local inconsistency cue would help readers.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The central claim is well supported; the two major points are disclosure/clarification rather than correctness risks. Fit for a solid CV/forensics venue is good. I would not block acceptance over the inpainting weakness if the authors add the requested bounds and the γ held-out note."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"This is a solid engineering paper that delivers what it claims. The useful novelty is not inventing temporal volatility in frozen features (D3 and ReStraV already did that), but moving the measurement to patch tokens with an explicit one-step vs two-step roughness ratio (TTR), adding local spatial gradients of the feature motion field (LSMI), fusing them with a single fixed gamma, and then evaluating at the exact operating point services care about: Fake Recall at 0.1% FPR under real-only calibration and cross-real transfer.\n\nThe numbers are strong and consistent. On FakeParts, GenVideo, and ViF-Bench it posts 74%, 85%, and 85% Fake Recall at 0.1% FPR while D3 collapses and the best supervised baselines lag badly, especially on the 2024–2025 generators. Ablations show the patch-level shift is the main lift, LSMI helps at the strictest threshold, gamma=8 saturates cleanly, and the gain holds across transformers and CNNs. Code is public, overhead is negligible, and post-processing robustness is decent. The evaluation protocol itself is a real contribution; AUROC tables in the appendix confirm that aggregate metrics hide how poorly global methods behave at ultra-low FPR.\n\nSoft spots are real but proportionate. Inpainting is still weak (2.5% at 0.1% FPR) because the signal is spatially sparse; they rank second and disclose it. The core assumption—that TTR/LSMI remain generator-agnostic under one fixed gamma and real-only thresholds—is the right one to watch, but they already stress-test it with ViF-Bench, encoder sweeps, and cross-real FPR transfer that stays near target. No circularity, no hidden fitting, no derivation gap.\n\nThis is for people building or evaluating detectors for platforms and forensics, and for anyone tired of AUROC-only papers. It deserves a serious referee. I would accept it for review and would cite the protocol and the patch-level results.","headline":"Training-free patch-level detector that actually works at 0.1% FPR on full and partial fakes, with a service-aligned protocol that is more useful than AUROC.","tokens_in":45621,"tokens_out":531,"would_cite":true,"duration_ms":6828,"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":"A training-free detector finds AI-generated and partially edited videos by measuring rough patch trajectories and incoherent local motion in frozen vision features, at 0.1% false-positive rates on real video.","keywords":["AI-generated video detection","training-free detection","partial video manipulation","patch tokens","temporal roughness","motion incoherence","ultra-low FPR","real-only calibration"],"falsifier":"A new video generator (or a post-processing pipeline) that produces patch-token trajectories whose one-step and two-step lengths stay as well matched as real video, and whose feature-space motion fields remain spatially coherent, would drive Fake Recall at 0.1% FPR toward the levels of existing baselines and falsify the claim.","tokens_in":45696,"feed_emoji":"🎬","tokens_out":972,"duration_ms":9942,"temperature":0.7,"pith_summary":"Real-world video services cannot afford to reject authentic user uploads, so detectors must keep the false-positive rate on real video extremely low (around 0.1%). Standard scores such as AUROC hide how poorly many methods actually operate at that threshold. This paper introduces SPLIT, a training-free method that never sees a single synthetic example. It simply extracts patch tokens from a frozen vision encoder and looks for two statistical signatures of manipulation: trajectories that are rough rather than smooth (one-step path length exceeds the normalized two-step chord) and motion fields whose neighboring patches disagree. The two signals are multiplied with a fixed gamma correction and thresholded using only real video. On three complementary benchmarks covering partial edits, earlier generators, and the newest 2024–2025 systems, the method records the highest Fake Recall at the strictest operating point while remaining robust to blur, compression, and flips and adding almost no cost beyond the encoder forward pass. The authors also supply a service-aligned evaluation protocol—Fake Recall at fixed FPR with real-only calibration and cross-real threshold transfer—so that future detectors can be judged under the constraints that actually matter for deployment.","feed_headline":"Patch roughness spots AI video at 0.1% false-positive rate","feed_subtitle":"Training-free scores on frozen tokens beat supervised detectors on full fakes and partial edits","key_machinery":"Two-step Temporal Roughness (TTR) together with Local Spatial Motion Incoherence (LSMI), fused as Score = TTR^γ · LSMI (γ = 8). TTR compares accumulated consecutive-step displacements of each patch token against a normalized two-step chord; LSMI measures how much neighboring patches disagree in their temporal feature changes. The product, thresholded on real video alone, is the detector.","core_discovery":"Localized spatiotemporal incoherence measured on patch tokens of a frozen vision encoder—specifically the log-ratio of one-step versus two-step feature path lengths (TTR) and the average magnitude of spatial gradients of the feature-space motion field (LSMI)—yields a single score that separates both fully generated and sparsely edited video from real video at ultra-low false-positive rates after real-only threshold calibration, without any task-specific training.","pith_inferences":["The same patch-level roughness statistics may transfer to still-image partial edits if the two-step idea is replaced by multi-scale spatial path-length comparisons.","If future generators begin to explicitly regularize feature-space trajectories for smoothness, the method will force a measurable quality–detectability trade-off that can be tracked over time.","The real-only calibration protocol itself is a reusable template for any high-stakes media detector (audio, image, video) that must guarantee ultra-low false-positive rates without target-domain fakes."],"forward_implications":["Service operators can set a single real-only threshold and obtain high recall on both fully synthetic and partially edited uploads without collecting or labeling generator-specific training data.","Partial-edit benchmarks such as FakeParts become practical stress tests rather than near-failure modes, because the detector never collapses the frame into a global embedding.","Cross-real threshold transfer can replace ad-hoc AUROC reporting as the primary metric for video-manipulation detectors that must operate under fixed false-positive budgets.","Any frozen vision encoder that yields patch tokens can be swapped in; the scoring logic itself adds negligible overhead and requires no fine-tuning."],"fun_headline_variants":["Patch roughness catches AI and partial-edit video at 0.1% FPR","Frozen-token TTR and LSMI flag generated video without training","Spatial incoherence plus temporal roughness detects deepfakes training-free","SPLIT scores patch trajectories to separate real from AI video at low FPR","One- and two-step path lengths on patches reveal AI fakes and edits"],"cache_read_input_tokens":32896,"weakest_assumption_plain":"That the excess of one-step over two-step patch roughness and of local motion-field gradients remains a stable, generator-agnostic cue once a single fixed gamma and a real-only threshold are chosen.","fun_headline_variants_meta":{"raw":{"variants":["Patch roughness catches AI and partial-edit video at 0.1% FPR","Frozen-token TTR and LSMI flag generated video without training","Spatial incoherence plus temporal roughness detects deepfakes training-free","SPLIT scores patch trajectories to separate real from AI video at low FPR","One- and two-step path lengths on patches reveal AI fakes and edits"]},"model":"grok-4.5","effort":"low","cost_usd":0.004966,"raw_usage":{"total_tokens":1441,"prompt_tokens":824,"num_sources_used":0,"completion_tokens":101,"cost_in_usd_ticks":49660000,"prompt_tokens_details":{"text_tokens":824,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":516,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":824,"tokens_out":101,"duration_ms":4586,"temperature":1.0,"reasoning_tokens":516,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T06:21:53.653708+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"A new video generator (or a post-processing pipeline) that produces patch-token trajectories whose one-step and two-step lengths stay as well matched as real video, and whose feature-space motion fields remain spatially coherent, would drive Fake Recall at 0.1% FPR toward the levels of existing baselines and falsify the claim.","supporting_citations":[],"review_version":1}