Abstract: The Longmenshan (LMS) fault zone is located at the junction of the eastern Tibetan Plateau and the Sichuan Basin and is of great significance for studying regional tectonics and earthquake hazards. Although regional velocity models are available for the LMS fault zone, high-resolution velocity models are lacking. Therefore, a dense array of 240 short-period seismometers was deployed around the central segment of the LMS fault zone for approximately 30 days to monitor earthquakes and characterize fine structures of the fault zone. Considering the large quantity of observed seismic data, the data processing workflow consisted of deep learning-based automatic earthquake detection, phase arrival picking, and association. Compared with the earthquake catalog released by the China Earthquake Administration, many more earthquakes were detected by the dense array. Double-difference seismic tomography was adopted to determine V_p, V_s, and V_p/V_s models as well as earthquake locations. The checkerboard test showed that the velocity models have spatial resolutions of approximately 5 km in the horizontal directions and 2 km at depth. To the west of the Yingxiu–Beichuan Fault (YBF), the Precambrian Pengguan complex, where most of earthquakes occurred, is characterized by high velocity and low V_p/V_s values. In comparison, to the east of the YBF, the Upper Paleozoic to Jurassic sediments, where few earthquakes occurred, show low velocity and high V_p/V_s values. Our results suggest that the earthquake activity in the LMS fault zone is controlled by the strength of the rock compositions. When the high-resolution velocity models were combined with the relocated earthquakes, we were also able to delineate the fault geometry for different faults in the LMS fault zone.