Image quality assessment (IQA), especially perceptual image IQA, has been researched for years. Automatic human perceptual quality is hard because when people evaluate picture quality, there are many unpredictable factors affecting their final scores. Therefore, it is necessary to build a large dataset for model training. It is costly to build such a dataset as both human annotation collection and data processing steps are tedious. In that case, there are not many datasets ever tested on large databases. In this paper, we will focus on the process of building the LIVE-FB large-scale Social Picture Quality Database, a deep IQA model evaluation, and new IQA model proposal.