High-throughput imaging techniques acquire large amounts of images efficiently. These images contain rich biological information including cellular processes. A common method to analyze them is to encode them into quantitative representation vectors. Generally, there are two ways to extract cell biological information into representations, hand-crafted and machine-learning. Although representations obtained from machine learning models often demonstrate commendable reconstruction performance, they lack biological interpretability. In contrast, hand-crafted representations have clear biological meanings, making them easily interpretable. However, the capability of hand-crafted representations to generate realistic images remains uncertain. In this work, we propose a CellProfiler to image (CP2Image) model capable of directly generating realistic cell images from CellProfiler representations. The proposed model is demonstrated to be robust to different architectures, including ResNet, InceptionNet and Transformer. We also show that the biological information is well preserved during the generation process. The changes in certain CellProfiler features will reflect the corresponding changes in the generated single-cell images. In addition, the CP2Image model can generate conditional phenotypes, which will ultimately help diagnostics and drug screening.