01 / 04
CVPR 2025 HighlightGenerative Photography: Scene-Consistent Camera Control for Realistic Text-to-Image Synthesis
Control the camera, preserve the world.
Scene-consistent camera control for realistic text-to-image synthesis.
ProblemExisting generative models do not understand the physical meaning of camera intrinsics or preserve scene consistency.
IdeaDimensional lifting disentangles scene and camera; differentiated parameter injection enables fine-grained control.
ResultControl bokeh, focal length, shutter speed, and color temperature while preserving scene identity.
02 / 04
ICLR 2026NewtonGen: Physics-Consistent and Controllable Text-to-Video Generation via Neural Newtonian Dynamics
Give video generation a sense of physics.
Physics-consistent and controllable text-to-video generation.
ProblemVideo models lack physical consistency and parameter-level controllability.
IdeaLearn Neural Newtonian Dynamics, then guide video generation with explicit physical states.
ResultDiverse, physically consistent, and controllable video generation.
04 / 04
Under ReviewOptiWorld: Optimal Control for Video World Generation under Physical Constraints
Plan a better future, then render it.
OptiWorld makes video world models choose better futures: safer, smoother, more efficient, and physically constrained motions before rendering.
ProblemCurrent video generation lacks proactive understanding and control, producing results that may be unsafe or inefficient.
IdeaIntroduce classical optimal control before video rendering.
ResultSafer, more efficient goal-conditioned I2V and video refinement.
