AffectFlow-DINO Uses Rectified Flow for Uncertainty-Aware Affect Estimation
July 13, 2026
AffectFlow-DINO implements a conditional rectified-flow head on a frozen DINOv3 ViT-S/16 backbone to model facial behavior ambiguity. The generative approach enables uncertainty-aware predictions, improving valence-arousal CCC-V scores by 0.058.
HOW THIS AFFECTS YOU
●
researcherYou can model one-to-many prediction distributions in facial expression tasks using Monte Carlo sampling.