6G wireless communication systems pose unprecedented challenges to the accuracy, efficiency, and environmental adaptability of channel models. Starting from fundamental electromagnetic theories, this presentation extends to the statistical modeling of highly dynamic and complex scenarios, and finally explores the new paradigm of AI-empowered channel generation. First, targeting cutting-edge 6G technologies such as holographic metasurfaces, a continuous-aperture channel discretization method based on Gauss-Legendre quadrature is proposed; this improves channel reconstruction accuracy by multiple orders of magnitude compared to classical methods, enabling multi-scenario small-scale statistical modeling. Second, addressing the 6G requirements for space-air-ground-sea integration, we have conducted land-to-ship channel measurements and modeling, and propose statistical models for large- and small-scale fading as well as sea-wave-induced fixed-point fading. We also utilize the Gini index to quantify channel sparsity and accurately characterize beyond-line-of-sight propagation laws. Third, to resolve the environmental adaptability issues of statistical modeling, a voxelized radiance field channel prediction method is proposed, which improves training efficiency by approximately 47 times compared to traditional neural radiance fields and reduces data requirements by 80%. Furthermore, a channel prediction framework based on vision-language models and panoramic RGB-D is constructed, realizing the intelligent generation of channel information such as multipath power, delay, and angles of arrival.