Tat-Jen Cham

Realistic human-object interactions in complex 3D scenes are now achievable by decoupling scene feasibility from motion synthesis, leveraging separate supervision signals.

Unlearning PII from VLMs is now feasible without sacrificing retained knowledge by directly manipulating attention-value representations.

Unreliable teacher predictions, when framed as negative samples, unlock robust real-world adverse weather removal without synthetic data.

Text-to-video models can now generate videos that actually respect spatial relationships, thanks to a new geometry-based reward function that beats VLM-based alternatives.