Bubble separation and size detecting algorithms are developed in recent years for their promise applications, which include bubble column reactor monitoring, cell counting in vivo, oil droplet characterization in petroleum, etc. In this work, we proposed an architecture called bubble generative adversarial networks (BubGAN) to bridge the gap between the image processing algorithm development and benchmark in bubbly flow measurement. The BubGAN is trained initially on a labeled bubble dataset with ten thousand real bubble images. By learning the distribution of these bubbles, the BubGAN generates a database with million synthetic bubbles. Using this database, BubGAN can then assemble genuine bubbly flow images and provide detailed bubble information with labels on the synthetic images. The BubGAN can serve as a useful tool to benchmark the existing image processing algorithms and to help to guide the future development of bubble detecting algorithms.