Hyesong Choi

Standard pixel augmentations frequently corrupt delicate vision-language alignment, but injecting diffusion-style isotropic noise directly into embedding spaces breaks through the longstanding performance ceiling of stacked CutMix, Mixup, and RandAug recipes.

AI struggles to fake realistic eye contact: inconsistencies in social gaze reveal synthetically generated images with surprising reliability.

Stop relying on global image-text alignment scores for vision-language pretraining data curation – a phrase-level sensitivity signal reveals a 50% data subset that substantially boosts compositional generalization.