Face Forgery Detection (FFD) is crucial to preserve the authenticity and trustworthiness of digital visual content. Most existing FFD methods perform well under intra-domain setups but do not generalize well to unseen domains. Recently, Large Vision-Language Models (LVLMs) have shown unprecedented generality in the computer vision community, making it a competitive potential choice for FFD. However, LVLMs lack specific domain knowledge and have a weak understanding of localized details within forged faces, which limits their effectiveness in the FFD task. In this paper, we explore the utilization of LVLM to address the cross-domain FFD problem and propose ForgeryGPT, a novel FFD approach based on LVLM. To integrate FFD knowledge into the LVLM, we employ a decoder to match the text and visual features to obtain localization results of the forgery artifacts. For better generalization to unseen domains, we design a prompt learner to leverage the powerful reasoning capabilities of large language models for better optimization. Moreover, the prompt learner enables our method to preserve the LVLM’s inherent capabilities and support multi-turn dialogues. We conduct comprehensive experiments on two commonly used cross-domain FFD benchmarks. The experimental results show that our proposed method significantly surpasses the exsiting state-of-the-art approaches. Our code is available at https://github.com/woody-panda/ForgeryGPT .