We present f VDB, a novel GPU-optimized framework for deep learning on large-scale 3D data. f VDB provides a complete set of differentiable primitives to build deep learning architectures for common tasks in 3D learning such as convolution, pooling, attention, ray-tracing, meshing, etc. f VDB simultaneously provides a much larger feature set (primitives and operators) than established frameworks with no loss in efficiency: our operators match or exceed the performance of other frameworks with narrower scope. Furthermore, f VDB can process datasets with much larger footprint and spatial resolution than prior works, while providing a competitive memory footprint on small inputs. To achieve this combination of versatility and performance, f VDB relies on a single novel VDB index grid acceleration structure paired with several key innovations including GPU accelerated sparse grid construction, convolution using tensorcores, fast ray tracing kernels using a Hierarchical Digital Differential Analyzer algorithm (HDDA), and jagged tensors. Our framework is fully integrated with PyTorch enabling interoperability with existing pipelines, and we demonstrate its effectiveness on a number of representative tasks such as large-scale point-cloud segmentation, high resolution 3D generative modeling, unbounded scale Neural Radiance Fields, and large-scale point cloud reconstruction.

Humans and animals can control their bodies to generate a wide range of motions via low-dimensional action signals representing high-level goals. As such, human bodies and faces are prime examples of active objects, which can affect their shape via an internal actuation mechanism. This paper explores the following proposition: given a training set of example poses of an active deformable object, can we learn a low-dimensional control space that could reproduce the training set and generalize to new poses? In contrast to popular machine learning methods for dimensionality reduction such as auto-encoders, we model our active objects in a physics-based way. We utilize a differentiable, quasistatic, physics-based simulation layer and combine it with a decoder-type neural network. Our differentiable physics layer naturally fits into deep learning frameworks and allows the decoder network to learn actuations that reach the desired poses after physics-based simulation. In contrast to modeling approaches where users build anatomical models from first principles, medical literature or medical imaging, we do not presume knowledge of the underlying musculature, but learn the structure and control of the actuation mechanism directly from the input data. We present a training paradigm and several scalability-oriented enhancements that allow us to train effectively while accommodating high-resolution volumetric models, with as many as a quarter million simulation elements. The prime demonstration of the efficacy of our example-driven modeling framework targets facial animation, where we train on a collection of input expressions while generalizing to unseen poses, drive detailed facial animation from sparse motion capture input, and facilitate expression sculpting via direct manipulation.

We present a multi-species model for the simulation of gravity driven landslides and debris flows with porous sand and water interactions. We use continuum mixture theory to describe individual phases where each species individually obeys conservation of mass and momentum and they are coupled through a momentum exchange term. Water is modeled as a weakly compressible fluid and sand is modeled with an elastoplastic law whose cohesion varies with water saturation. We use a two-grid Material Point Method to discretize the governing equations. The momentum exchange term in the mixture theory is relatively stiff and we use semi-implicit time stepping to avoid associated small time steps. Our semi-implicit treatment is explicit in plasticity and preserves symmetry of force linearizations. We develop a novel regularization of the elastic part of the sand constitutive model that better mimics plasticity during the implicit solve to prevent numerical cohesion artifacts that would otherwise have occurred. Lastly, we develop an improved return mapping for sand plasticity that prevents volume gain artifacts in the traditional Drucker-Prager model.

We simulate sand dynamics using an elastoplastic, continuum assumption. We demonstrate that the Drucker-Prager plastic flow model combined with a Hencky-strain-based hyperelasticity accurately recreates a wide range of visual sand phenomena with moderate computational expense. We use the Material Point Method (MPM) to discretize the governing equations for its natural treatment of contact, topological change and history dependent constitutive relations. The Drucker-Prager model naturally represents the frictional relation between shear and normal stresses through a yield stress criterion. We develop a stress projection algorithm used for enforcing this condition with a non-associative flow rule that works naturally with both implicit and explicit time integration. We demonstrate the efficacy of our approach on examples undergoing large deformation, collisions and topological changes necessary for producing modern visual effects.