AdaFocus: Instruction-Aware Dynamic Visual Token Compression for Efficient Multi-modal Understanding.

Zihang Ye, Xianzhong Liu

Browse the full ICPR paper archive.