Chao Liang

We propose methods to improve visual recognition in noisy, long-tailed datasets.

CapHuman generates realistic portraits from a single image, showing diverse versions of individuals.

A method to improve deepfake detection by reducing data biases.

IDRetracor helps trace original identities from manipulated faces, addressing deepfake challenges.

FriendsQA dataset improves video understanding by answering complex questions from Friends episodes.