Exactly. The AI shows an abstraction of an abstraction (to the AI its all just bits and patterns) but the thing that's immediately fascinating about the painting is that it is echoing someones experience in the flesh of a wild pig so long ago. The stuff about "humans ability to express abstraction" is very much in second place for me.
the replies to this comment are some of the least brain-intensive pieces of information you could ever come up with. do you think generative AI has not been trained on images of pigs?
it sounds to me like people don't like/are sceptical about generative AI, not that they're so expert in the area that they can give a valid opinion on whether this is possible or useful.
part of the technique that diffusion models use is literally to take a vague image of something and then use training data to build it into something more clear. this whole comment chain is so confidently wrong it's unbelievable
Something more clear is totally different than reconstructing.
If an artists rendition of what it might have looked like would be useful, then genAI may be cheaper for that task. But we wouldn't call an artist's rendition of what it might have looked like "reconstructing".
And you couldn't be much more wrong about how much I like and use generative AI.
I speculate your comment has been downvoted because you haven’t explained why you’d “like to see generative AI try to reconstruct the full image”.
What would be the purpose of that reconstruction? To help visualize the most likely original scene using a contemporary-biased lens? Why would we want that?
Imagine you found a box of macaron sandwich cookie fragments, and your model only knows about Oreos. Would such a reconstruction have value? Could it also apply inappropriate bias?
On the other hand, the right models can be critical. Look at how researchers created an image of black holes. [0] This is a reconstruction, and it is an average, and it relies on physics models; but those assumptions are in some ways more transparent than generative AI.
