The training data is presumably full of examples of people using the pattern to indicate transparency (and explaining that they do so — like the input for 50!),

The training data is presumably full of examples of people using the pattern to indicate transparency (and explaining that they do so — like the input for 50!), and much less of people actually creating such images (if the training data even preserves the alpha channel in the first place). I think a bigger problem is the "artifacts" you describe (worse than that sounds to me).