Yu Yuan · Research overview

Towards physics-consistent and controllable generative worlds.

Generative models should model worlds—not just pixels.

Navigate with ← →, Space, or swipe · M for map · F for full screen

The core research question

How can we build physics-consistent and controllable generative worlds?

Consistent across camera, time, and space. Controllable through conditions, goals, and constraints.

Cameraphysics-consistent image formation
Timecontrollable physical dynamics
Spaceconsistent 4D world completion
Actioncontrol under goals and constraints
01 / 04
Generative Photography · CVPR 2025 Highlight

Control the camera, preserve the world.

Control focal length, bokeh, shutter speed, and color temperature while preserving scene identity.

ProblemCamera prompts do not obey optics.
IdeaDimensional lifting disentangles scene and camera; differentiated parameter injection enables fine-grained control.
ResultControllable effects with stronger scene consistency.
Generative Photography camera-control pipeline
02 / 04
NewtonGen · ICLR 2026

Give video generation a sense of physics.

Learn Neural Newtonian Dynamics, then guide video generation with explicit physical states.

ProblemVideo models imitate motion without understanding it.
IdeaLinear neural ODEs plus a residual model.
ResultDiverse motion with explicit parameter control.
NewtonGen framework
03 / 04
SeeU · CVPR 2026

See beyond the observed time and space.

Lift sparse 2D observations into a continuous 4D world across unseen times and viewpoints.

ProblemObservations are sparse in both time and viewpoint.
IdeaModel continuous physical equations of motion in a 4D world.
ResultTemporal, spatial, and edited unseen-world generation.
SeeU continuous 4D world overview
04 / 04
OptiWorld · Under Review

Plan a better future, then render it.

Optimize goal-directed trajectories under physical constraints before rendering the future.

ProblemPlausible futures can still be unsafe or inefficient.
IdeaUnify constraints on a trajectory manifold, then actively perceive and plan the optimal path.
ResultGoal-conditioned and counterfactual world generation.
OptiWorld optimal-control overview