High throughput phenotyping (HTP) is an emerging frontier field across many basic and applied plant science disciplines. RGB imaging is most widely used in HTP to extract image-based phenotypes such as pixel volume or projected area. These image-based phenotypes are further used to derive plant physical parameters including plant fresh biomass, plant dry biomass, water use efficiency etc. In this paper, we investigated the robustness of regression models to predict fresh biomass of maize plants from image-based phenotypes. Data used in this study were from three different experiments. Data were grouped into five datasets, two for model development and three for independent model validation. Three image-derived phenotypes were investigated: BioVolume, Projected.Area.1, and Projected.Area.2. Models were assessed with R2, Bias, and RMSEP (Root Mean Squared Error of Prediction). The results showed that almost all models were validated with high R2 values, indicating that these digital phenotypes can be useful to rank plant biomass on a relative basis. However, in many occasions when accurate prediction of plant biomass is needed, it is important for researchers to know that models that relate image-based phenotypes to plant biomass should be carefully constructed. Our results show that the range of plant size and the genotypic diversity of the calibration sets in relation to the validation sets have large impact on the model accuracy. Large maize plants cause systematic bias as they grow toward the top-view camera. Excluding top-view images from modeling can there benefit modeling for the experiments involving large maize plants.