Saved in:
Bibliographic Details
Main Authors: Wu, Yanlin, Yuan, Xiaogang, An, Dezhi
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2511.09184
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866912704390758400
author Wu, Yanlin
Yuan, Xiaogang
An, Dezhi
author_facet Wu, Yanlin
Yuan, Xiaogang
An, Dezhi
contents AI-generated video has advanced rapidly and poses serious challenges to content security and forensic analysis. Existing detectors rely mainly on pixel-level visual cues and generalize poorly to unseen generators. We propose DBINDS, a diffusion-model-inversion based detector that analyzes latent-space dynamics rather than pixels. We find that initial noise sequences recovered by diffusion inversion differ systematically between real and generated videos. Building on this, DBINDS forms an Initial Noise Difference Sequence (INDS) and extracts multi-domain, multi-scale features. With feature optimization and a LightGBM classifier tuned by Bayesian search, DBINDS (trained on a single generator) achieves strong cross-generator performance on GenVidBench, demonstrating good generalization and robustness in limited-data settings.
format Preprint
id arxiv_https___arxiv_org_abs_2511_09184
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle DBINDS -- Can Initial Noise from Diffusion Model Inversion Help Reveal AI-Generated Videos?
Wu, Yanlin
Yuan, Xiaogang
An, Dezhi
Computer Vision and Pattern Recognition
AI-generated video has advanced rapidly and poses serious challenges to content security and forensic analysis. Existing detectors rely mainly on pixel-level visual cues and generalize poorly to unseen generators. We propose DBINDS, a diffusion-model-inversion based detector that analyzes latent-space dynamics rather than pixels. We find that initial noise sequences recovered by diffusion inversion differ systematically between real and generated videos. Building on this, DBINDS forms an Initial Noise Difference Sequence (INDS) and extracts multi-domain, multi-scale features. With feature optimization and a LightGBM classifier tuned by Bayesian search, DBINDS (trained on a single generator) achieves strong cross-generator performance on GenVidBench, demonstrating good generalization and robustness in limited-data settings.
title DBINDS -- Can Initial Noise from Diffusion Model Inversion Help Reveal AI-Generated Videos?
topic Computer Vision and Pattern Recognition
url https://arxiv.org/abs/2511.09184