zlacker

an interesting use case for this once the compute is there is to basically allow for ai powered digital zoom-out. it could work by instructing the user to take several pictures around the target, and then you take regular pictures of your subject.

then, as you like, you can do an "ai zoom out" to get zoomed out pictures, no longer constrained by your lens or distance.

I imagine this to be included relatively soon, just like how panaromas were once a niche thing that became much easier to do with some good ui/ux. pretty much any modern phone can do them without having to struggle with lining up photos and what not.

one thing that does greatly concern me about the demo/site is that they have "authentic" and "recover" as terms. the result here is not authentic nor has anything been "recovered." it's an illusion at best. I personally don't like how they portray the new image as being equivalent as if the lens framed it in the original picture. it's not, as they show themselves in the later portion (near the end) with the text sign. seriously irresponsible framing (pun intended) to what's otherwise very cool tech.

>>endisn+(OP)
Agreed, the "Reference-Driven Generation" part is totally fine, but "authentic" is overselling it.

>>endisn+(OP)
Nice idea. Might not need multiple pics given Google’s image dataset and ability to recognise what you’re looking at.

Give that a couple generations. “You were at location X and didn’t take a pic. We generated you some selfies, choose one that you like.”

>>richar+W8
If they wanted to, they could find real pictures of you, taken by other Google users who were there at the same time.

I don't know if that's more or less creepy than the AI stuff...

>>endisn+(OP)
This already exists in Stable Diffusion and others, called outpainting.