How hard would it be to automatically derive such maps from textures? After all, they are already drawn to visually imply that with shading, so wouldn't it be possible to e.g. train a model to extract that information, just as our eye does?