DeepSDF is a novel approach for representing 3D shapes using continuous Signed Distance Functions (SDFs). Unlike traditional methods that discretize SDFs into regular grids, DeepSDF learns a generative model to produce a continuous field. This continuous representation allows for high-quality shape representation, interpolation, and completion from partial and noisy 3D input data. The method uses a feed-forward neural network to predict the SDF value at any query point, with the zero-level set of the learned function implicitly encoding the shape's boundary. DeepSDF can represent an entire class of shapes and achieves state-of-the-art performance in shape reconstruction and completion while significantly reducing model size compared to previous methods. The key contributions include the formulation of generative shape-conditioned 3D modeling, a learning method based on a probabilistic auto-decoder, and the demonstration of these techniques in shape modeling and completion. DeepSDF produces high-quality continuous surfaces with complex topologies and provides accurate surface normals. The method is evaluated on various datasets, showing superior performance in representing known and unknown shapes, shape completion from partial observations, and shape interpolation in the latent space.DeepSDF is a novel approach for representing 3D shapes using continuous Signed Distance Functions (SDFs). Unlike traditional methods that discretize SDFs into regular grids, DeepSDF learns a generative model to produce a continuous field. This continuous representation allows for high-quality shape representation, interpolation, and completion from partial and noisy 3D input data. The method uses a feed-forward neural network to predict the SDF value at any query point, with the zero-level set of the learned function implicitly encoding the shape's boundary. DeepSDF can represent an entire class of shapes and achieves state-of-the-art performance in shape reconstruction and completion while significantly reducing model size compared to previous methods. The key contributions include the formulation of generative shape-conditioned 3D modeling, a learning method based on a probabilistic auto-decoder, and the demonstration of these techniques in shape modeling and completion. DeepSDF produces high-quality continuous surfaces with complex topologies and provides accurate surface normals. The method is evaluated on various datasets, showing superior performance in representing known and unknown shapes, shape completion from partial observations, and shape interpolation in the latent space.