Don't take it personal. Commenters on HN are famous for dismissing successful ideas (remember Dropbox?).

I have one question: you mentioned that the training data was 100GB. Was it the same resolution as what is output by the model (ignoring supersampling)?