The Metaverse is a virtual world based on numerous technologies, which enables users to interact socially in a persistent online 3-D virtual environment. To generate high-level imaginary environments, extremely low latency data transmission and learning-based sensor data analysis are required. With the development of 5G techniques, processing and learning methods, both the transmission delay and high-quality scene generation have been significantly improved in meta-applications. However, many Metaverse devices are battery-powered, and local processes and learning are still too costly. To address this issue, in this paper, by taking full architectural advantage of 5G networks, we propose a novel dynamic split learning scheme for enabled Metaverse systems. In our proposed scheme, each neural network is split into two segments, and the upper segment is stored at the base station (BS) side. Thus, between two segments, multiple pathways are featured, each with distinct compression ratios, accompanied by a gating mechanism that intelligently guides the selection of paths for each input data. This design excels in adapting to diverse Metaverse applications and network conditions, enhancing both the learning and computing phases of split models. Simulation results underscore the efficacy of our proposed scheme, revealing that it does not impede the convergence of split learning models. Furthermore, the scheme demonstrates notable performance gains in terms of communication overhead, prediction accuracy, and adaptability to resource constraints.