Generative models such as GANs and Diffusion Models show promise in network traffic synthesis, yet often fail to jointly deliver high fidelity, strict protocol compliance, and fine-grained parameter control. We propose CascadeGen, a two-stage hybrid framework that first employs a parameter-aware conditional GAN to generate accurate flow-level statistics, then uses a conditional Diffusion Model to refine them into high-fidelity packet-level nPrint representations. A post-processing module further enforces protocol correctness to ensure practical usability. On public datasets, CascadeGen significantly surpasses single-stage baselines, achieving up to 40.2% JSD reduction for statistical similarity, a 20.6% FID reduction for packet-level fidelity, and a 6.2% point increase in PCR, while also demonstrating over 90.1% MRE reduction for precise control of key attack parameters. CascadeGen provides an advanced framework for generating high-quality, customizable synthetic traffic in the field of network security.