Explainable Forensics of Manipulated Segments in Untrimmed Long Videos

Yue Feng^1*, Jingjing Li^2*, Qijia Lu^1*, Wei Ji^3‡*, Jingrou Zhang¹, Fei Shen⁴, Xiao Li¹, Yizhen Jia¹, Qiang Chen^3‡, Limin Wang⁵, Wentong Li^1†, Jie Qin^1†

^*Equal Contribution ^†Corresponding Author ^‡Project Lead

¹MoE Key Laboratory of Brain-Machine Intelligence Technology, College of Artificial Intelligence, Nanjing University of Aeronautics and Astronautics ²Dalian University of Technology ³Independent Scholar ⁴National University of Singapore ⁵Nanjing University

🎉 The dataset and code will be open-sourced before July 1, 2026. Stay tuned!

Video Demos

Task Definition

Illustration of the TASLE task, focusing on localizing sparsely embedded AI-generated segments within long untrimmed videos.

Dataset Overview

In the following video demo, we showcase some examples from the TASLE dataset.

Annotation Examples

Visual examples of TASLE annotations, illustrating authenticity labels, temporal boundaries, and multi-level rationales.

Proposed Framework

Overview of MSLoc. The proposed coarse-to-fine framework integrates boundary-sensitive proposal generation with MLLM-based refinement to achieve precise localization and explanation.

Qualitative Results

Qualitative comparison of MSLoc with other models, highlighting its ability to accurately detect authenticity and provide interpretable analysis in complex scenes.