Ivy-Fake: A Unified Explainable Framework and Benchmark for Image and Video AIGC Detection

Ivy-Fake is a novel, unified, and large-scale dataset specifically designed for explainable multimodal AI-Generated Content (AIGC) detection. It is the first large-scale benchmark for this purpose, covering both images and videos. The project also introduces Ivy-XDetector, a unified vision-language model that achieves state-of-the-art performance in detecting and explaining AI-generated images and videos.

The rapid advancement of AIGC has led to hyper-realistic synthetic media, raising significant concerns about authenticity and misinformation. Current detection methods often operate as black-box classifiers, lacking interpretability and unified support for different modalities. Ivy-Fake addresses these limitations by providing a comprehensive benchmark and an explainable detection framework.

Key Features

• Unified Multimodal Dataset: Ivy-Fake contains over 150,000 richly annotated training samples (images and videos) and 18,700 evaluation examples.
• Explainable Annotations: Each sample is accompanied by detailed natural-language reasoning, going beyond simple binary labels to explain detected spatial and temporal artifacts.
• IVY-XDETECTOR Model: A unified vision-language architecture that performs state-of-the-art detection and provides human-readable explanations for both image and video AIGC.
• Focus on Real-World Scenarios: The dataset includes diverse content generated by various cutting-edge architectures (GANs, Diffusion models, Transformers) and authentic content from real-world contexts.
• Publicly Available: The Ivy-Fake dataset is publicly accessible on Hugging Face Datasets.

Project Overview

The Ivy-Fake framework (illustrated below) conducts an in-depth analysis of temporal and spatial artifacts to enable explainable detection of AI-generated content.

Caption: Overview of the IVY-FAKE framework: By conducting in-depth analysis of temporal and spatial artifacts, the framework enables explainable detection of AI-generated content.

The IVY-XDETECTOR model follows a three-stage progressive training framework:

1. General Video Understanding: Equipping the model with fundamental video comprehension.
2. AIGC Detection Fine-tuning: Specializing the model for binary AIGC discrimination (real vs. fake).
3. Joint Optimization for Detection and Explainability: Enabling the model to generate high-quality explanations while maintaining detection accuracy.

Getting Started

Dataset Access

The Ivy-Fake dataset can be accessed on Hugging Face: https://huggingface.co/datasets/AI-Safeguard/Ivy-Fake

Code

The code for the IVY-XDETECTOR and related experiments can be found in this repository: https://github.com/Pi3AI/IvyFake (Further instructions on setting up the environment, running the code, and training/evaluating models will be provided here or in a separate CONTRIBUTING.md or documentation files.)

Citation

If you use Ivy-Fake or IVY-XDETECTOR in your research, please cite our paper:

@article{jiang2025ivyfake,
  title     = {Ivy-Fake: A Unified Explainable Framework and Benchmark for Image and Video AIGC Detection},
  author    = {Jiang, Changjiang and Zhang, Wayne and Zhang, Zhonghao and Yu, Fengchang and Peng, Wei},
  year      = {2025}
}