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

Advancements in AIGC, driven by models like DALL-E, Imagen, Stable Diffusion, and SORA, have led to highly realistic synthetic images and videos, posing challenges to content authenticity and public trust. Most current AIGC detection methods are binary classifiers with limited interpretability and often lack support for diverse generators or modalities. While Multimodal Large Language Models (MLLMs) show promise for explainable detection, existing benchmarks are inadequate, often lacking video data or sufficient annotation depth.

To overcome these limitations, we introduce IVY-FAKE, a comprehensive benchmark with diverse multimodal data and rich, explainable annotations. Building on this, our IVY-XDETECTOR model excels at identifying and explaining spatial and temporal generative artifacts in both images and videos. Our key contributions are a unified vision-language detector and the first large-scale benchmark for explainable multimodal AIGC detection.

The IVY-FAKE dataset is designed for explainable multimodal AIGC detection, containing 100K+ training images, around 5,000 total test samples. It features diverse content (animals, objects, DeepFakes, etc.) and sources (GANs, Diffusion models, Transformers) and is kept current by collecting new AIGC content.

Video data was sourced from public benchmarks like GenVideo and LOKI, and web-crawled platforms, including outputs from models like SORA and Stable Video Diffusion. Image data was similarly collected from public datasets like FakeClue and WildFake, and web sources, covering GANs and Diffusion models. A stratified sampling strategy ensures balanced representation.

Explainable annotations were generated using Gemini 2.5 Pro with a knowledge distillation process and a structured template requiring reasoning before conclusion (<think>...</think><conclusion>...</conclusion>). Gemini was provided with ground-truth labels and a prompt like "This {file_type} is {label}. Explain the reason.". Explanations are categorized into Spatial Features (8 sub-dimensions) and Temporal Features (4 sub-dimensions, video-only). IVY-FAKE uniquely integrates explainable annotations across both image and video modalities, unlike prior datasets which are often limited in scope or modality.

Table 1: Example of IVY-FAKE annotations.

This section describes the extensive experiments conducted to evaluate the detection and explanation capabilities of the IVY-XDETECTOR. The model was tested on classification tasks (distinguishing real from fake content) for both images and videos, using metrics like accuracy and F1-score. Reasoning abilities were assessed by comparing model-generated explanations with reference annotations using ROUGE-L scores and an LLM-as-a-judge approach (evaluating completeness, relevance, detail, and explanation quality). The proposed method demonstrated superior performance on various benchmarks for both image and video content classification and produced more transparent reasoning compared to several leading multimodal models.