All your megapixels are now dead

6 years ago

Wait am I to understand that Google is ahead of TikTok?

as if there was ever any doubt 😎

Which clearly implies that you have a very cool trendy dance division at Google that I have yet to discover!

Yeah, Gigapixel AI astounds me with with what it does with thumbnail photos

One day, our robot masters will look back at this and say “01010100 01101000 01100101 00100000 01100110 01101111 01101111 01101100 01110011 00100000 01110111 01100101 01110010 01100101 00100000 01110011 01101111 00100000 01100011 01101100 01101111 01110011 01100101 00100110”

The point stands but does anyone with good conscience or photography background really think the 3rd image is "better" than the original? To me the original is still much better.

It's not really the storage of data though, it's the creation of interpolated data. For imagery it's an "artistic" interpretation at best. It's akin to making a watercolour of an oil painting.

@tomstandage Good question. Before AI, image compression was about removing what you might not notice (or what could be tokenized) to make a much smaller filer with a decent amount of verity. This is through-the-looking-glass stuff and could mean a lot of untrustworthy compression results.

One of the reasons I was excited about the xh2s was because the sensor was “only” 26 MP. It’s fine for what I shoot and the rare times I need it larger I’ve got Topaz Gigapixel.

This might be interesting for some niches where cost of storage or transmission is much higher than computational cost. Would be interesting to see the final photo after 5 or 10 or 100 cycles of downsizing and restoring with AI🤔

in ML terms most of the relevant information lies on a low dimensional manifold embedded within high dimensions. Aim of a good AI algorithm is to discover that low dimensional manifold

"Enhance 224 176" The future is now.

Interesting. While the AI-generated one might not be "better" in terms of accuracy, it does appear to produce a more finished "look" for the images.

@dogfood Oh, that reminds me of the Xerox bug that made a lot of digitized data from paper unreliable. Dangerous because a lot of people will rely on it without seeing the change.

and who is this person?

"Plausible" replaces "true", which is in tune with the probabilistic inference engine that powers successful AI (it turns out the brain already does that, so in fact we don't look at a picture, or at anything, we fantasize it). Leads to "Mise en abyme":⬇️

It's interesting and potentially useful, a synthetic recreation. But in art the process is important and usually in photography I want the dirt, grain and natural texture so it won't replace it :)

Yeah - not a fan. I wouldn't say it looks better than the original, either, but an "enhanced" somewhat unrealistic version.

The AI version has blurry spots.