TruthLens is a novel DeepFake detection framework that goes beyond the limitations of traditional binary classification (real vs. fake) to determine if an image is real or fake and provides detailed text-based reasoning for that prediction. Its hybrid design combines the global contextual understanding of multi-modal large-scale language models, such as PaliGemma2, with the local feature extraction of vision-only models, such as DINOv2, to effectively handle both face-manipulation DeepFakes and AI-generated content. It can even answer questions about fine details such as eyes, nose, and mouth, and has shown 2-14% better detection accuracy and explainability than existing state-of-the-art methods on a variety of datasets. It generalizes well to existing and novel manipulation techniques.
