With the rise of the concept of the metaverse, virtual scene generation technology has become the key to building an immersive digital world. However, traditional virtual scene generation methods have problems such as low efficiency, lack of realism, and lack of dynamic interactivity, which makes it difficult to meet the high requirements of the metaverse for scene diversity and real-time performance. This paper uses GAN to generate realistic scene textures and structures; secondly, the scene elements are globally modeled and optimized through Transformer; finally, reinforcement learning is combined to achieve dynamic interaction of the scene. Through experiments on multiple virtual scene datasets, the algorithm is superior to traditional methods in generation efficiency, visual quality, and interactive response speed, effectively improving the generation quality and user experience of the metaverse virtual scene. The research results show that through the evaluation of scene fidelity, high-configuration hardware improves the baseline fidelity from 85.3 to 89.5%. Furthermore, the hybrid enhancement strategy achieves additional improvements of 1.8%, 3.6%, and 0.8% under different configurations, respectively, and the final fidelity reaches 90.3%. The diversity index shows a synergistic optimization effect: in the case of low configuration and no enhancement strategy, the diversity is 81.2%, while when high-configuration hardware is combined with the hybrid enhancement strategy, the index is significantly improved to 87.5%. This paper theoretically enriches the application of deep learning in the field of virtual scene generation, and provides important technical support for the practical application of Metaverse technology, which has high theoretical significance and practical value.