We present a refined 3D crustal model beneath SE Tibet from ambient noise adjoint tomography. Different from ray-theory-based tomography, adjoint tomography in this study incorporates a spectral-element method (SEM) and takes empirical Green's functions (EGFs) of Rayleigh waves from ambient noise interferometry as the direct observation. The frequency-dependent traveltime misfits between SEM synthetic Green's functions and EGFs are minimized with a preconditioned conjugate gradient method, meanwhile the 3D model gets improved iteratively utilizing 3D finite-frequency kernels. The new model shows 3 - 6% shear wave speed increasing beneath the western Sichuan Basin (SCB) (depth>15 km) and the central Chuan-Dian Block (CDB), and 6 - 12% shear wave speed reduction in the mid-lower crust beneath the northern and the southern CDB. The inferred spatial pattern of low wave speed zones, consistent with possible partial melt, suggests more complex and disconnected geometry than the pervasive narrow zone from the channel flow models.