Convolutional neural networks are able to learn realistic image priors from numerous training samples in low-level image generation and restoration [66]. We show that, for high-level image recognition tasks, we can further recon struct “realistic” images of each category by leveraging intrinsic Batch Normalization (BN) statistics without any training data. Inspired by the popular VAE/GAN method s, we regard the zero-shot optimization process of synthet ic images as generative modeling to match the distribution of BN statistics. The generated images serve as a calibra tion set for the following zero-shot network quantizations. Our method meets the needs for quantizing models based on sensitive information, e.g., due to privacy concerns, no data is available. Extensive experiments on benchmark datasets show that, with the help of generated data, our ap proach consistently outperforms existing data-free quanti zation methods.