VLForgery introduces a comprehensive framework that extends deepfake detection capabilities by leveraging multimodal LLMs to not only detect synthetic faces but also locate forgeries and identify their source.

• Addresses high-quality face synthesis from diffusion models that evade current detectors
• Provides fine-grained analysis beyond binary detection (real/fake)
• Delivers visual localization of manipulated facial regions
• Attributes forgeries to specific generation methods

This research significantly enhances digital security by providing forensic tools that can explain the cause of forgeries, critical for identity verification systems and combating misinformation in an era of increasingly sophisticated AI-generated content.

VLForgery Face Triad: Detection, Localization and Attribution via Multimodal Large Language Models