Automatic propagation of real-world parent seismic images to efficiently generate a collection of realistic synthetic child training images to train a model for accurate automatic seismic interpretation. A 3D structural model in a present-day geological space (e.g., GB) depicting subsurface locations of particles (e.g., in region B) may be transformed by a 3D coordinate space transformation (e.g., uvtB) to a depositional space (e.g., G*B) depicting past depositional locations of those particles (e.g., corresponding depositional region B). A real-world parent image depicting subsurface locations of particles (e.g., in region A) may be transformed, via a forward transformation (e.g., uvtA), to a depositional seismic image in the depositional space of the three-dimensional structural model (e.g., G*A=G*B). A reverse transformation (e.g., uutB−1) may transform the depositional seismic image from the depositional space into synthetic child training images in the present-day geological space (e.g., GB) for training the model.