Featured project · Roblox · 2026

Roblox Reality

Video models as an AI renderer for interactive 3D worlds.

The game engine provides world structure and game logic; a video model transforms its output into richer visuals. Creator control guides the appearance, style, and freedom of the generative renderer.

Generating realistic urban crowd videos from coarse 3D simulations, with control over scene layout, camera motion, human trajectories, and appearance.

Generate coherent 3D assets with user-specified semantic parts, ready for animation and scripted behaviors.

Turning text into expressive, animatable 3D character heads by combining 3D generation with video diffusion.

Infer material properties from observed motion, then re-simulate objects under new conditions with a shared neural material model.

Fast, localized 3D mesh editing from text and rough masks, while preserving regions outside the edit.

A large-scale video diffusion model that can generate 4D video grids in a feed-forward manner.

Reconstructing detailed 3D meshes from multi-view images with improved geometry and texture.

Dense, per-pixel 3D tracking from monocular videos.

A foundation-model approach to 3D intelligence, connecting shape generation, scene understanding, and reasoning.

The first photorealistic text-to-4D scene generation pipeline, with the use of video diffusion models.

High-fidelity 3D scene generation from text.

A high-fidelity text-to-3D generation model with pretrained text-to-image models as prior.

Feed-forward generation of textured meshes from text, using amortized optimization with 2D diffusion priors.

Reconstructing dynamic scenes and synthesizing unseen views from a single video using RGB-D diffusion priors.

Diffusion Probabilistic Fields (ICLR 2023)

A diffusion model that learns distributions over continuous functions defined over metric spaces as fields.

Controllable radiance fields for dynamic face synthesis (3DV 2022)

A Controllable Radiance Field that can generate dynamic faces and human bodies with user control.

EMIXER: End-to-end Multimodal X-ray Generation via Self-supervision (MLHC 2022)

A multi-modal X-ray generative model that is trained in a self-supervised way.

AMICO: Amodal Instance Composition (BMVC 2021)

An image composition model that can blend imperfect objects onto a target image.

Enjoy Your Editing: Controllable GANs for Image Editing via Latent Space Navigation (ICLR 2021)

A latent-space image editing work using pretrained Generative Adversarial Networks (GANs).

FMRI data augmentation via synthesis (ISBI 2019)

Synthesizing fMRI brain data with generative adversarial networks for data augmentation.

Synthetic Power Analyses: Empirical Evaluation and Application to Cognitive Neuroimaging (Asilomar 2019)

A methodological approach to improve statistical analysis in cognitive neuroimaging using generated datasets.

Conditional structure generation through graph variational generative adversarial nets (NeurIPS 2019)

A variational generative model specifically designed for graphs.