It will be very useful to construct a set of notebooks showing how to train new models and use existing one (including viz of the results). My suggestion would be to build a model on ACS 814W data, akin to what is done in the original paper. The resultant models can be added to the repo (and docs/README) updated accordingly.