AffectFlow-DINO: Uncertainty-Aware Multi-Task Affect Estimation via Conditional Rectified Flow

AffectFlow-DINO targets the 11th ABAW challenge with a multi-task model built on a frozen DINOv3 ViT-S/16 backbone. Its conditional rectified-flow head learns a distribution of affect outputs rather than a single deterministic estimate, supporting Monte Carlo one-to-many predictions. The system jointly predicts continuous valence-arousal, eight facial-expression classes, and twelve Action Units from static face images. Reported ablations show a +0.058 gain in valence CCC from flow decoding. Post-hoc threshold calibration improves a severely imbalanced Fear class from 3.8% to 33.1%. The final system reports P_MTL=1.177 versus the official baseline’s 0.45.