Dalton Omens

I'm a Research Engineer at Epic Games, working on the next generation of AI-driven virtual characters and worlds. I recieved my PhD in Computer Science from the Stanford University AI Lab (SAIL), advised by Ron Fedkiw. I received my B.S. in Electrical Engineering and Computer Science at UC Berkeley in 2020.

I have worked with research teams at Epic Games, SONY, Unity, and Blizzard on cutting-edge computer graphics and AI initiatives. I have also interned at NVIDIA.

Email  /  CV  /  Bio  /  Scholar  /  LinkedIn  /  Github  /  Bluesky

Research

My current research centers upon AI for virtual characters and worlds: How do we improve world engines, linear content generation, and virtual characters with AI tools? How do we ground multimodal generative AI for consistency, efficiency, and usefulness?

More broadly, I continue to aim towards democratizing artistic pipelines, by leveraging cutting-edge tech to enable traditionally prohibitive workflows such as character simulation/creation, content generation, emergent gameplay systems, and animation/retargeting.

Using carefully captured calibration expressions can improve the semantic accuracy of facial retargeting. Implicit differentiation can be applied to black-box or non-differentiable trackers to fine-tune retargeting results for an even better result.

To simulate loose-fitting clothing, separating physically simulated dynamic modes from neurally simulated quasistatic modes aids generalization when paired with a well-crafted coarse physics model such as rope chains.

Obtaining geometry and texture from a couple selfies to create a personalized animation rig. A Simon Says process is used to match the user's expressions.

A differentiable physics engine that supports Lagrangian dynamics and hard contact constraints for articulated rigid body simulation, offering analytic gradients through features typically only available in non-differentiable simulators.

A method using convolutional neural networks for approximating the mesh deformation of complex facial rigs. Fast evaluation allows for interactive inverse kinematics and real-time performance capture.

Education