Attention Registers vs Global Attention

Click nodes to inspect. Toggle modes to view information routing and memory utilization.

INPUT TOKENS MEDIATION & STORAGE DOWNSTREAM HEADS MEMORY OVERHEAD (N² VS N) 92% FRAME_01 FRAME_02 FRAME_03 CAMERA_TOK REG_BANK [4] CAMERA HEAD DENSE HEAD TEXT ALIGN HEAD
Click any token or processing block to inspect its role and attention behavior.