I'd like to see generative AI try to reconstruct the full image
I think you may be struggling with the word "works". and "exampled". and perhaps even "trained"
vague image + entire corpus of human imagery = ?
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
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.
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.
[0]: https://youtu.be/Ol_SB5Zfv-Y?si=O6CL-kjDcgBB0bWm , 4-minute overview