We have presented FlexDataset, a groundbreaking framework for generating high-fidelity synthetic datasets tailored to diverse perceptual tasks such as salient object detection, depth estimation, and generic segmentation. FlexDataset pioneers a composition-to-data (C2D) generation paradigm, enabling the creation of complex, multi-instance and multi-category (MIMC) scenes that closely resemble real-world environments. Our Versatile Annotation Generation (VAG) Plan A enhances annotation synthesis efficiency while maintaining exceptional quality. Comprehensive experiments demonstrate that FlexDataset surpasses existing text-to-data methods, underscoring its potential to transform dataset creation by reducing the need for manual annotation and accelerating the development of robust deep networks.