Galaxy spectral energy distributions (SEDs) remain among the most challenging yet informative quantities to reproduce in simulations due to the large and complex mixture of physical processes that shape the radiation output of a galaxy. With the increasing number of surveys utilizing broad-band colours as part of their target selection criteria, the production of realistic SEDs in simulations is necessary for assisting in survey design and interpretation of observations. The recent success in reproducing the observed luminosity functions (LFs) from far-ultraviolet (UV) to far-infrared (IR), using the state-of-the-art semi-analytic model SHARK and the SED generator PROSPECT, represents a critical step towards better galaxy colour predictions. We show that with SHARK and PROSPECT we can closely reproduce the optical colour distributions observed in the panchromatic Galaxy And Mass Assembly (GAMA) survey. The treatment of feedback, star formation, central-satellite interactions, and radiation reprocessing by dust are critical for this achievement. The first three processes create a bimodal distribution, while dust attenuation defines the location and shape of the blue and red populations. While a naive comparison between observation and simulations displays the known issue of overquenching of satellite galaxies, the introduction of empirically motivated observational errors and classification from the same group finder used in GAMA greatly reduces this tension. The introduction of random reassignment of similar to 15 per cent of centrals/satellites as satellites/centrals on the simulation classification closely resembles the outcome of the group finder, providing a computationally less intensive method to compare simulations with observations.