Kyungjae Lee

Achieving fine-grained semantic alignment in text-to-video generation is now possible with a model that explicitly verifies every prompt condition against visual evidence.

Escape the jagged landscape of trajectory optimization: natural functional gradients smooth the path to feasible robotic motion in complex environments.

Forget finicky hyperparameters and fragile smoothing schedules: this new probabilistic smoothing method provably finds the global optimum with greater robustness.

Forget dense pose targets: sparse taxonomy guidance unlocks dexterous manipulation with surprising generalization and controllability.