Existing data-driven methods for garment animation, usually driven by linear skinning, although effective on tight garments, do not handle loose-fitting garments with complex deformations well. To address these limitations, we develop a garment generative model based on deformation decomposition to efficiently simulate loose garment deformation without directly using linear skinning. Specifically, we learn a garment generative space with the proposed generative model, where we decouple the latent representation into unposed deformed garments and dynamic offsets during the decoding stage. With explicit garment deformations decomposition, our generative model is able to generate complex pose-driven deformations on canonical garment shapes. Furthermore, we learn to transfer the body motions and previous state of the garment to the latent space to regenerate dynamic results. In addition, we introduce a detail enhancement module in an adversarial training setup to learn high-frequency wrinkles. We demonstrate our method outperforms state-of-the-art data-driven alternatives through extensive experiments and show qualitative and quantitative analysis of results.