World Action Models are becoming more popular in robotics, as they learn to predict the world jointly with learning how to act on it. However, these world predictions are usually purely based on reconstructing color images from video. This is limiting, because color is far from