A single research agenda
at increasing levels of abstraction.

We make video look better at the pixel level. By exploiting temporal structure across frames, we synthesize smoother motion, restore detail beyond a single frame's limits, stabilize shaky perspective, and minimize quality loss under bandwidth constraints. Each task reduces to one core problem — modeling motion and correspondence between frames precisely enough to act on it.

Connecting threadAll tasks reduce to precise modeling of inter-frame motion and correspondence.

We extract semantic meaning from temporal visual data — understanding what changes in a video and why. From action recognition to large-scale video foundation models, we study representations that capture meaningful temporal structure rather than treating video as a sequence of independent images.

Connecting threadTemporal semantics is the bridge from pixels to meaning — understanding what changes, and why.

We reconstruct 3D scenes from 2D observations and render them from new viewpoints. Neural representations — radiance fields, Gaussian splatting — capture geometry and appearance with high fidelity. From this foundation, we extend naturally into 4D, modeling how scenes evolve over time.

Connecting threadRecovering 3D structure from 2D observations — naturally extending into time when scenes move.

We build generative systems for the visual world — from single-image synthesis with diffusion, through action-conditioned and physics-aware video generation, and into world foundation models that let robots and embodied agents anticipate the consequences of their actions. Across these directions, the aim is the same: visual intelligence — perception, prediction, and simulation — for systems that act in the world.

Connecting threadGeneration, prediction, and action-conditioning share one foundation — and one purpose: equipping agents to anticipate the visual consequences of action.