I don't know about that. I recall reading something from a while ago where AI was used to upscale a license plate. The AI was trained specifically on low resolution license plates with known...
I don't know about that. I recall reading something from a while ago where AI was used to upscale a license plate. The AI was trained specifically on low resolution license plates with known numbers. It could then reverse a blurry plate with an unknown number successfully. It seems like something that could be given a high % of certainty.
Edit:
Here'sa paper on it. I don't think that's what I saw originally.
On second thought, you wouldn't be so interested in this as evidence for a conviction, but as support for a warrant.
And here is what it looks like when you upscale an image of obama https://video-images.vice.com/test-uploads/articles/5ef216e9aea1330097d9ec64/lede/1592924226088-obama.jpeg These upscales are a...
These upscales are a work of fiction. A plausible reality but not the actual truth. Good enough for printing on the side of a bus for an advert but not enough to gain more info than you had originally and trusting that data to be correct.
While I'm sure it could be used in that way, I'm more concerned about the opposite happening, as with these copiers that changed numbers in spreadsheets. While that particular flaw wasn't...
While I'm sure it could be used in that way, I'm more concerned about the opposite happening, as with these copiers that changed numbers in spreadsheets. While that particular flaw wasn't AI-related, I worry that AI producing such subtle issues may go unnoticed.
Newer versions of Photoshop have some incredible machine learning baked in. Content aware crop and fill are effectively magic, and even when they don't give totally perfect results they get the...
Newer versions of Photoshop have some incredible machine learning baked in. Content aware crop and fill are effectively magic, and even when they don't give totally perfect results they get the job 90% done in two clicks.
It seems like a lot of research has really come to fruition in the last few years - as a non-expert user, it feels like there are more meaningful changes for me between the 2018 and 2021 versions than there were in pretty much the decade prior.
I’ve been playing around with it for a bit. It generally works well if the input has low JPEG artifacting, otherwise it really amplifies the artifacts. Sometimes high noise images (from high ISO,...
I’ve been playing around with it for a bit. It generally works well if the input has low JPEG artifacting, otherwise it really amplifies the artifacts. Sometimes high noise images (from high ISO, usually) can produce weird colour artifacts. Images with a lot of crisp details in focus also tend to do better than softer images.
This is a good example of what it can pull off in less than ideal circumstances. It's not earth shattering, but it's a good deal better than old school upscaling. Original Enhanced
This is a good example of what it can pull off in less than ideal circumstances. It's not earth shattering, but it's a good deal better than old school upscaling.
I'm just amused that you activate it with "Enhance".
It's mad to think it was a scifi joke for ages and now is a legitimate thing. Not to the extent it often was seen in TV but still more so than before.
Well, legitimate fakery anyway. You wouldn’t want to use this on a photo being used as evidence.
I don't know about that. I recall reading something from a while ago where AI was used to upscale a license plate. The AI was trained specifically on low resolution license plates with known numbers. It could then reverse a blurry plate with an unknown number successfully. It seems like something that could be given a high % of certainty.
Edit:
Here's a paper on it. I don't think that's what I saw originally.
On second thought, you wouldn't be so interested in this as evidence for a conviction, but as support for a warrant.
And here is what it looks like when you upscale an image of obama https://video-images.vice.com/test-uploads/articles/5ef216e9aea1330097d9ec64/lede/1592924226088-obama.jpeg
These upscales are a work of fiction. A plausible reality but not the actual truth. Good enough for printing on the side of a bus for an advert but not enough to gain more info than you had originally and trusting that data to be correct.
While I'm sure it could be used in that way, I'm more concerned about the opposite happening, as with these copiers that changed numbers in spreadsheets. While that particular flaw wasn't AI-related, I worry that AI producing such subtle issues may go unnoticed.
Newer versions of Photoshop have some incredible machine learning baked in. Content aware crop and fill are effectively magic, and even when they don't give totally perfect results they get the job 90% done in two clicks.
It seems like a lot of research has really come to fruition in the last few years - as a non-expert user, it feels like there are more meaningful changes for me between the 2018 and 2021 versions than there were in pretty much the decade prior.
I’ve been playing around with it for a bit. It generally works well if the input has low JPEG artifacting, otherwise it really amplifies the artifacts. Sometimes high noise images (from high ISO, usually) can produce weird colour artifacts. Images with a lot of crisp details in focus also tend to do better than softer images.
This is a good example of what it can pull off in less than ideal circumstances. It's not earth shattering, but it's a good deal better than old school upscaling.
Original
Enhanced
Note that's afaik it's pretty much a copy of Topaz Gigapixel, so it's not really something innovative.