This paper presents FLAME, a new 3D facial model that combines shape, pose, and expression variations to enable realistic facial animation. FLAME is trained from over 33,000 3D scans, including 3800 scans of human heads and 4D sequences from the D3DFACS dataset. The model is designed to be compatible with existing graphics software and is easy to fit to data. FLAME uses a linear shape space trained from 3800 scans of human heads, combined with an articulated jaw, neck, and eyeballs, pose-dependent corrective blendshapes, and additional global expression. The model is trained to minimize 3D reconstruction error and is available for research purposes. FLAME is significantly more accurate and expressive than existing face models such as the FaceWarehouse model and the Basel Face Model. The model is trained from data to capture realistic 3D face details and enable animation. The model is evaluated on three datasets, including the BU-3DFE database and the Beeler et al. sequence. The results show that FLAME is significantly more expressive than existing models and can be used to fit 2D image data and for expression transfer. The model is trained using an iterative optimization approach that minimizes the reconstruction error of the training data. The pose parameters are trained first, followed by the expression parameters, to avoid expression overfitting. The shape parameters are trained by computing the template and shape blendshapes for the registrations in the shape dataset. The model is evaluated quantitatively on the D3DFACS database and our self-captured sequences, showing that 60% of the vertices have a median distance less than 0.2mm, and 90% are closer than 0.5mm. The model is also evaluated qualitatively, showing that it can capture subtle facial motions such as eye blinks. The model is compared to the Basel Face Model and FaceWarehouse model, showing that FLAME is significantly more expressive. The model is available for research purposes and includes software to animate and use the model. The temporal registration of the D3DFACS dataset is also made publicly available for research purposes.This paper presents FLAME, a new 3D facial model that combines shape, pose, and expression variations to enable realistic facial animation. FLAME is trained from over 33,000 3D scans, including 3800 scans of human heads and 4D sequences from the D3DFACS dataset. The model is designed to be compatible with existing graphics software and is easy to fit to data. FLAME uses a linear shape space trained from 3800 scans of human heads, combined with an articulated jaw, neck, and eyeballs, pose-dependent corrective blendshapes, and additional global expression. The model is trained to minimize 3D reconstruction error and is available for research purposes. FLAME is significantly more accurate and expressive than existing face models such as the FaceWarehouse model and the Basel Face Model. The model is trained from data to capture realistic 3D face details and enable animation. The model is evaluated on three datasets, including the BU-3DFE database and the Beeler et al. sequence. The results show that FLAME is significantly more expressive than existing models and can be used to fit 2D image data and for expression transfer. The model is trained using an iterative optimization approach that minimizes the reconstruction error of the training data. The pose parameters are trained first, followed by the expression parameters, to avoid expression overfitting. The shape parameters are trained by computing the template and shape blendshapes for the registrations in the shape dataset. The model is evaluated quantitatively on the D3DFACS database and our self-captured sequences, showing that 60% of the vertices have a median distance less than 0.2mm, and 90% are closer than 0.5mm. The model is also evaluated qualitatively, showing that it can capture subtle facial motions such as eye blinks. The model is compared to the Basel Face Model and FaceWarehouse model, showing that FLAME is significantly more expressive. The model is available for research purposes and includes software to animate and use the model. The temporal registration of the D3DFACS dataset is also made publicly available for research purposes.