It’s… tricky.  The particular model architecture used to predict the images below has around 1.2 million trainable parameters, but I’ve plans to decrease that complexity and greatly improve accuracy.