You might recall my first such experiment. It was not completely successful, but a good learning experience. I’m here now with a different attempt. Here are comparison images of a token I've got coming in a recent swap. The first image in each sequence is the dealer's picture, unaltered. The second is AI-enhanced by the Google Gemini AI bot, which was instructed to edit out the pixels where the white prongs overlapped the edge of the token - and only those pixels. I haven't yet done a really close examination, but this looks like a pretty good result to me. Not only are the pesky prongs gone, but the image is a more usable size for my purposes: 1024 x 1024 instead of the original 500 x 500. (My usual standard is 800 x 800.) 1A- obverse, unaltered 1B- obverse, AI-enhanced 2A- reverse, unaltered 2B- reverse- AI-enhanced
Upon my first casual inspection, I'm leaning toward the opinion that this set of AI images might actually be usable? I think that instructing the bot to edit only the pixels where the prongs overlapped the edge of the token - and nothing else -might have been the all-important prompt here. Are there any tiny details I'm overlooking? The AI images look pretty faithful to the original, as far as I can tell. See anything that's "off"? Any differences in color, texture, marks, etc.?
That is nicely done! The pesky prongs is one reason why I'm not keen using slabbed coin photos for type set. This is one set of photos that I would be happy to use!
OK, I’m noticing what seems to be a very subtle difference in color (the blues/greens?) but perhaps not enough of one to concern me.
Which means our vaulted AI can be used to enhance any image of any coin. Are you listening, Ebay? ("Hey, this doesn't look like the coin I bought!")
Oh, that particular cat has long been out of the bag, I’m afraid. Those with nefarious intent have much more powerful weapons in their arsenal, now.
My hatred of white NGC prongs intruding into images is pathological- so much so that I’ve been teased for it. Especially on dinky coins like gold dollars or California fractionals, where the prongs swallow unacceptable swaths of the visible surfaces.
Yep, that's why I've restricted my coin collecting to just a few (very few) auction sites and what I find on the sidewalk.
Looks like a pretty good job to me. There was a small thread-like object on the slab above one of the obverse prongs that wasn't removed (it's above the "O" in "STORE"). You programmed it not to remove anything besides the prongs so that should be expected.
I agree, at a quick review these look "unmodified" except the removal of the prongs. That is the key part I think. Just yesterday I did an AI experiment with ChatGPT to overlay 1807 and 1809 Bust Half dollars to compare the changes that happened in 1809. (eagles long neck reduced, changes to drapery on the bust, etc). Here is what I found: (1) First, AI made the high resolution 1807 "better" by subtly redesigning it. Very subtle. It did this I think to avoid copyright issues with PCGS and to clean up the image (remove toning for example). (2) AI also made the high resolution 1809 "better" by subtly redesigning it. Very subtle. (3) Compared the "better" version of 1807 to "better" version of 1809 and overlaid some very nice graphics. (4) Looked beautiful, professional, and very "believable", AND totally wrong. It would likely fool most coin collectors that didn't specialize in Capped Bust Half dollars. I pointed this out to the AI, and it agreed "fixing" the images beforehand would lead to a totally erroneous comparison. And it proceeded with a new image, and did it again! How does this relate to what you are doing? Well, I think you likely avoided this problem, but it seems all too easy for AI to want to "fix" / "improve" images beyond the scope of someone's prompt. In my case, the result wasn't just misleading, the conclusion of how the 1807 design differed from the 1809 design was totally wrong.
Good catch. Would’ve taken me approximately 3.5 years to notice that. It is, as your observation suggests, actually a point of confidence for me that the AI was following instructions.
