Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
AI-Generated Video Detection via Perceptual Straightening
Authors: Christian Internò, Robert Geirhos, Markus Olhofer, Sunny Liu, Barbara Hammer, David Klindt
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | A lightweight classifier achieves state-of-the-art detection performance (e.g., 97.17% accuracy and 98.63% AUROC on the Vid Pro M benchmark [3]), substantially outperforming existing imageand video-based methods. ... We demonstrate through extensive experiments on diverse benchmarks (Vid Pro M [3], Gen Vid Bench [17], and Physics-IQ [26]) and models, that Re Stra V improve detection accuracy that often surpasses state-of-the-art (So TA) methods. |
| Researcher Affiliation | Collaboration | Christian Internò1,3 Robert Geirhos2 Markus Olhofer3 Sunny Liu4 Barbara Hammer1 David Klindt4 1 Bielefeld University 2 Google Deep Mind 3Honda Research Institute EU 4Cold Spring Harbor Laboratory |
| Pseudocode | No | The paper describes the method and equations in text and figures, but does not include a distinct section or block explicitly labeled as 'Pseudocode' or 'Algorithm'. |
| Open Source Code | Yes | Code: https://github.com/Christian Interno/Re Stra V |
| Open Datasets | Yes | The Vid Pro M dataset is offered for non-commercial research purposes under CC BY-NC 4.0 license. ... The Gen Vid Bench dataset and its associated code are under CC BY-NC-SA 4.0 license. ... The Physics-IQ dataset is available under the Apache License 2.0. ... DVSC2023 is under the CC BY-SA 4.0 llicense. |
| Dataset Splits | Yes | We use the dataset from Section 5 and apply a stratified 50/50 train/test split. Class priors are identical, and each subset is balanced among five AI models (Pika [46], Video Craft2 [62], Text2Video-Zero [63], Model Scope [3] and Sora [64]). |
| Hardware Specification | Yes | All experiments presented in this paper were conducted on a system equipped with NVIDIA RTX2080 GPUs, each with 8GB of VRAM. |
| Software Dependencies | No | The paper mentions DINOv2 Vi T-S/14 model and the scenedetect library, but does not provide specific version numbers for these or other software dependencies like Python, PyTorch, or TensorFlow. |
| Experiment Setup | Yes | For classification, we construct a feature vector y per video by combining direct signals and aggregated statistics from these trajectories. Specifically, y concatenates seven distance values [d1, d2, . . . , d7] and six curvature values: [θ 1, θ 2, . . . , θ 6]; and four statistical descriptors (mean, variance, minimum, maximum) for both {di} and {θ i }. This results in a final feature vector y R21. ... We consider only off-the-shelf models: logistic regression (LR), Gaussian Naive Bayes (GNB), random forest (RF; 400 trees, depth 6), gradient boosting (GB; 200 rounds, learning rate 0.1), RBF-kernel SVM (calibrated by Platt scaling), and a two-layer MLP (64 32). |