FaceForensics++ is a dataset of facial forgeries designed to enable researchers to train deep-learning-based approaches for detecting manipulated facial images. The dataset includes manipulations created using four state-of-the-art methods: Face2Face, FaceSwap, DeepFakes, and NeuralTextures. The paper proposes an automated benchmark for facial manipulation detection, which is based on these four methods and considers realistic scenarios with random compression and dimensions. The benchmark includes a hidden test set and a database of over 1.8 million manipulated images, making it significantly larger than comparable datasets. The authors perform a thorough analysis of data-driven forgery detectors and show that using additional domain-specific knowledge improves detection accuracy, even in the presence of strong compression, outperforming human observers. The paper also introduces a novel large-scale dataset of manipulated facial imagery, composed of more than 1.8 million images from 1,000 videos, and evaluates state-of-the-art hand-crafted and learned forgery detectors in various scenarios. The contributions of the paper include an automated benchmark for facial manipulation detection, a large-scale dataset, and a state-of-the-art forgery detection method tailored to facial manipulations.FaceForensics++ is a dataset of facial forgeries designed to enable researchers to train deep-learning-based approaches for detecting manipulated facial images. The dataset includes manipulations created using four state-of-the-art methods: Face2Face, FaceSwap, DeepFakes, and NeuralTextures. The paper proposes an automated benchmark for facial manipulation detection, which is based on these four methods and considers realistic scenarios with random compression and dimensions. The benchmark includes a hidden test set and a database of over 1.8 million manipulated images, making it significantly larger than comparable datasets. The authors perform a thorough analysis of data-driven forgery detectors and show that using additional domain-specific knowledge improves detection accuracy, even in the presence of strong compression, outperforming human observers. The paper also introduces a novel large-scale dataset of manipulated facial imagery, composed of more than 1.8 million images from 1,000 videos, and evaluates state-of-the-art hand-crafted and learned forgery detectors in various scenarios. The contributions of the paper include an automated benchmark for facial manipulation detection, a large-scale dataset, and a state-of-the-art forgery detection method tailored to facial manipulations.