>>5176
> I'd also like to 'help' the user by saying 'A has higher resolution' and so on, although we may have to revisit it if there are many false positives.
AFAIK, there will be.
What you can usually do with reasonable reliability is eliminate the images that have been more lossily compressed. You can do this by running them through some of the better modern codecs (BPG, WEBP or something like that).
A decision between an original image and one with a more or less obvious effect or watermark or whatever applied and all these things is probably not something you can realistically program at this point.
I don't think there is any realistic method other than what many *boorus do: Let registered human users tag image versions as better as others and kinda share that information while letting an editor resolve contradictory reports.