Official PyTorch implementation of the NeurIPS 2025 paper "Seeing What Matters: Generalizable AI-generated Video Detection with Forensic-Oriented Augmentation". Riccardo Corvi, Davide Cozzolino, Ekta Prashnani, Shalini De Mello, Koki Nagano, Luisa Verdoliva.

• 2026-04-18: release of the augmentation code.
• 2026-01-06: release of the code to compute the power spectra and the test set.
• 2025-11-28: demo release.

In this work, first, we study different generative architectures, searching and identifying discriminative features that are unbiased, robust to impairments, and shared across models. Then, we introduce a novel forensic-oriented data augmentation strategy based on the wavelet decomposition and replace specific frequency-related bands to drive the model to exploit more relevant forensic cues. Our novel training paradigm improves the generalizability of AI-generated video detectors, without the need for complex algorithms and large datasets that include multiple synthetic generators. To evaluate our approach, we train the detector using data from a single generative model and test it against videos produced by a wide range of other models. Despite its simplicity, our method achieves a significant accuracy improvement over state-of-the-art detectors and obtains excellent results even on very recent generative models, such as NOVA and FLUX.

You can find the code to run the method’s demo in the demo folder. For more details, check the accompanying README file.

You can find the code to compute the power spectra in the demo folder. For more details, check the accompanying README file.

You can find the code to download the test set in the dataset folder. For more details, check the accompanying README file.

You can find the code of the augmentation in the augmentation folder.

The license of the code can be found in the LICENSE.md file.