[ExI] Google AI low-resolution photo enhancemen
pharos at gmail.com
Sun Sep 12 09:30:36 UTC 2021
On Fri, 10 Sept 2021 at 21:51, Dave Sill via extropy-chat
<extropy-chat at lists.extropy.org> wrote:
> Exactly. To go from low-res to high-res without additional information requires making up new data. The Google photo enhancement AI is good at making up data that results in believable portraits--and it only does portraits at this point.
> None of the stories about this enhancement have compared the results of enhancing an image with the actual high-res version of the portrait. They all published the same Google-supplied images which were no doubt cherry-picked to make the software look good.
> Again, this technology will be fine for upscaling images for non law-enforcement purposes, but it in no way recovers data lost due to low resolution.
Another article from a photography website is also enthusiastic and
gives a bit more information about the processing involved.
You are obviously correct in saying that to start with a low-res photo
and create a high-res photo involves adding information that is not
present in the original. Nobody is disputing that. What Google is
claiming is that training their AI software on similar photos means
that they can create a high-res photo that is photo-realistic and very
likely to be a correct image.
Again obviously, the more information that is present in the original
low-res image then the more likely it is that the high-res result will
be correct. For example, a small birthmark or spot that is not shown
in the low-res image will not magically appear in the high-res image.
One interesting use might be enhancing old family photos.
More information about the extropy-chat