FaceForensics++ is a large-scale dataset of facial forgeries designed to enable the training of deep learning-based methods for detecting manipulated facial images. The dataset includes over 1.8 million manipulated images generated using four state-of-the-art methods: Face2Face, FaceSwap, DeepFakes, and NeuralTextures. These methods were applied to 1,000 pristine videos, and the resulting images were compressed and resized to simulate real-world conditions. The dataset is significantly larger than existing publicly available forgery datasets and includes a hidden test set for evaluation. The paper introduces an automated benchmark for facial manipulation detection, which includes a human baseline. The benchmark evaluates the performance of both human observers and automated systems in detecting manipulated facial images. The results show that automated systems, particularly those using deep learning, outperform human observers in detecting forgeries, even under challenging conditions such as high compression and low image quality. The paper also presents a comprehensive evaluation of state-of-the-art forgery detection methods, including both hand-crafted and learned features. The results demonstrate that deep learning-based approaches, such as XceptionNet, achieve higher accuracy in detecting forgeries compared to traditional methods. Additionally, the paper introduces a novel dataset and benchmark for facial forgery detection, which is publicly available for further research and development. The study highlights the importance of domain-specific knowledge in improving the accuracy of forgery detection. The proposed dataset and benchmark provide a standardized framework for evaluating and comparing different detection methods, facilitating further research in digital media forensics. The results show that automated systems can reliably detect facial forgeries, even in challenging scenarios, and that the proposed methods are effective in distinguishing between manipulated and genuine images. The paper concludes that the proposed dataset and benchmark are valuable resources for the research community in the field of digital media forensics.FaceForensics++ is a large-scale dataset of facial forgeries designed to enable the training of deep learning-based methods for detecting manipulated facial images. The dataset includes over 1.8 million manipulated images generated using four state-of-the-art methods: Face2Face, FaceSwap, DeepFakes, and NeuralTextures. These methods were applied to 1,000 pristine videos, and the resulting images were compressed and resized to simulate real-world conditions. The dataset is significantly larger than existing publicly available forgery datasets and includes a hidden test set for evaluation. The paper introduces an automated benchmark for facial manipulation detection, which includes a human baseline. The benchmark evaluates the performance of both human observers and automated systems in detecting manipulated facial images. The results show that automated systems, particularly those using deep learning, outperform human observers in detecting forgeries, even under challenging conditions such as high compression and low image quality. The paper also presents a comprehensive evaluation of state-of-the-art forgery detection methods, including both hand-crafted and learned features. The results demonstrate that deep learning-based approaches, such as XceptionNet, achieve higher accuracy in detecting forgeries compared to traditional methods. Additionally, the paper introduces a novel dataset and benchmark for facial forgery detection, which is publicly available for further research and development. The study highlights the importance of domain-specific knowledge in improving the accuracy of forgery detection. The proposed dataset and benchmark provide a standardized framework for evaluating and comparing different detection methods, facilitating further research in digital media forensics. The results show that automated systems can reliably detect facial forgeries, even in challenging scenarios, and that the proposed methods are effective in distinguishing between manipulated and genuine images. The paper concludes that the proposed dataset and benchmark are valuable resources for the research community in the field of digital media forensics.