Search papers, labs, and topics across Lattice.
Harbin Institute of Technology, Harbin, China
1
0
3
Reducing visual tokens doesn't always mean faster inference; a pre-vision strategy can significantly cut latency by bypassing preprocessing steps.
