Is this image compression tool good at images it was not trained on?
How bad does it get in those situations?
Is this training data fixed into the codex forever? Will there be slightly different image codexs that have different training data? That would be sort of hellish.