We propose DAA, a trainable module that enables real-time adaptation by amplifying feature-level discrepancies between known and unknown classes in TTD.