Alt
At the top is a screenshot from Wikimedia Commons showing an image that was updated to a larger size with the comment saying “Improved image”. Below it is the goose chasing meme with the goose twice asking “Where did the pixels come from?”.
At the top is a screenshot from Wikimedia Commons showing an image that was updated to a larger size with the comment saying “Improved image”. Below it is the goose chasing meme with the goose twice asking “Where did the pixels come from?”.
This seems like one of the few good use cases for AI, no?
I’ve never seen an AI upscaled image that didn’t look like shit. Even if it looked good it would still be artificial. We could go into the philosophy of weather or not a photo is an accurate snapshot of a moment in time, but think of it this way: let’s say we have a photo of you. Let’s also say you want to upscale that photo of you to some redicously high resolution for some reason. The AI upscaled image created is no longer a photo of you. It is an machine’s best guess at what you look like in Super 14K or whatever resolution. In essence, It’s inauthentic.
600KB vs 1800KB for no added information. Do you at least donate to the Wikimedia foundation so that they can pay for the hosting cost?
Why would a higher resolution even matter? You cant see extra details because those details are made up
AI upscaling is great for creative endeavors but not ideal for anything trying to be an objective source of truth about the world.
No.
The increased “detail” is entirely made up, based off whatever the AI model considers likely to be there based off similar images. AI isn’t somehow magically finding pixels that don’t exist, it’s effectively just guessing them.
I swear, were none of you people around for the LSD dogslugs of early image generation/style matching? Or the slightly newer “this person is not real” portrait generators that would merge hair and glasses, and often give multiple sets of eyes when glasses were at odd angles? This is effectively that with considerably more training data thrown at it.
It’s all made up. The AI isn’t taking another picture of the object with a higher resolution camera. It’s spreading out the existing pixels and doing a best guess to fill in the blanks. Maybe that’s fine for a family portait or something (I don’t agree, but you do you), but that’s definitively not OK for any sort of actual reference material.
Like that time when a upscaler turned Barack Obama into a white man.
Thanks for the link. I don’t know whether I should laugh or cry. It’s hilariously bad, and people embrace this humanity-destroying tech cheerfully.
How do you think image enhancement has always worked? Do you think AI tools are ignoring 30-years of Photoshop equations?
And no, we shouldn’t be fiddling with primary sources. Do you think Wikipedia is a primary source?
Upscaling can introduce errors which is not something you’d want in an encyclopedia
Only if it’s vetted by a professional, really. Upscaling sometimes introduces artifacts which you really wouldn’t want on something like this.
And keep the original because in 5 years someone is going to upload an “improved” version of the “improved” version