MicrostockGroup
Microstock Photography Forum - General => General Stock Discussion => Topic started by: melastmohican on October 24, 2023, 14:45
-
It is in the early stage of research but I think eventually every image out in the wild will be "poisoned" :-)
https://petapixel.com/2023/10/24/nightshade-data-poisoning-tool-helps-creatives-protect-art-from-ai/
https://gizmodo.com/nightshade-poisons-ai-art-generators-dall-e-1850951218
You can download this tool now if you are curious: https://glaze.cs.uchicago.edu/what-is-glaze.html
-
It is in the early stage of research but I think eventually every image out in the wild will be "poisoned" :-)
https://petapixel.com/2023/10/24/nightshade-data-poisoning-tool-helps-creatives-protect-art-from-ai/
https://gizmodo.com/nightshade-poisons-ai-art-generators-dall-e-1850951218
You can download this tool now if you are curious: https://glaze.cs.uchicago.edu/what-is-glaze.html
cute approach, but highly unlikely:
- billions of images already out there
- no way to poison images at agencies
- artists would have to poison thousands of their images
-
no way to poison images at agencies
Really, why? Will they sue us? lol
i think it's not even possible to detect cloaking / poisoning on an image
-
no way to poison images at agencies
Really, why? Will they sue us? lol
i think it's not even possible to detect cloaking / poisoning on an image
I think what he means is that you can't poison images you already have on the agencies. You'd have to delete and re-upload. Not practical.
-
no way to poison images at agencies
Really, why? Will they sue us? lol
i think it's not even possible to detect cloaking / poisoning on an image
I think what he means is that you can't poison images you already have on the agencies. You'd have to delete and re-upload. Not practical.
One possible scenario is that an agency could apply this to their whole database. For example, Getty is already in a lawsuit against an AI company for copyright infringement. If they protected their whole database, a contributor could go exclusive with them.
Just a thought.
(yes, I know, we used to be really angry with Getty for bad treatment of contributors. It would be funny if they ended up being our saviors ???)
-
I think what he means is that you can't poison images you already have on the agencies. You'd have to delete and re-upload. Not practical.
True, I got that wrong.
-
It is in the early stage of research but I think eventually every image out in the wild will be "poisoned" :-)
https://petapixel.com/2023/10/24/nightshade-data-poisoning-tool-helps-creatives-protect-art-from-ai/
https://gizmodo.com/nightshade-poisons-ai-art-generators-dall-e-1850951218
You can download this tool now if you are curious: https://glaze.cs.uchicago.edu/what-is-glaze.html
cute approach, but highly unlikely:
- billions of images already out there
- no way to poison images at agencies
- artists would have to poison thousands of their images
True but as new poisoned images are added, the dirty work will be done. It doesn't need to be 100%. Even 10% erroneous data could really mess up the machine learning.
-
I have bad news: I have tried it and it is useless.
Maybe if you just want to use your images online in small size it might work, but in full size the altering is extremely visible, even at a low glaze setting. It just looks like it makes the quality of the image bad.
There is a lot of fragmentation that gets added to one-colored areas, like this. And they are really all over the picture.
(https://i.postimg.cc/8kvG1g4R/001.jpg)
And here are some side by side comparisons of unglazed and glazed full size crops:
(https://i.postimg.cc/YqmvKPGZ/002.jpg) (https://i.postimg.cc/rm0DVZZD/003.jpg)
(https://i.postimg.cc/yxnxvqkr/005.jpg) (https://i.postimg.cc/qBZzhrHz/Untitled-1.jpg)
(https://i.postimg.cc/fR8p7MD0/006.jpg) (https://i.postimg.cc/6QGc7K3K/006a.jpg)
(https://i.postimg.cc/FRYxtQM8/008.jpg) (https://i.postimg.cc/Wbv08r1Q/008a.jpg)
The blurry parts are taken from the defocused backgrounds of images, that's where the glaze effect is the worst. Parts that are sharp and and in focus and are very small look okay-ish, but of course you rarely have a photo where everything is in focus and where you have no bigger areas with the same color. You can see that on the eye example. I cannot tell a difference with the dog fur around the eyes, it looks good, but the iris itself, where there is larger colored brown area looks weird.
Here is bigger sized example, but this one is already sized down 75% - and you can still notice the effect very much. First one unedited, second glazed.
(https://i.postimg.cc/9frx2TTL/00000.jpg) (https://i.postimg.cc/dVDRZF11/00000a.jpg)
There is no way any microstock agency would accept images of this quality.
These examples are from the recommended low glaze setting with little "protection". If you use the highest setting the images are completely butchered:
(https://i.postimg.cc/WzFFbNSL/0.jpg)
-
That is bad news - I wasn't sure if it could make a difference because of the various powerful interests who want to just take whatever they find useful, but with that kind of visible damage to an image it's useless for anything but social media stuff.
Thanks very much for taking the time to explore this and report on what you observed.
-
Thanks for the examples @firn
As far as I know, Nightshade hasn't been integrated into Glaze yet.
Glaze is designed for concealing your style (gives weird results when you use a cloaked image as prompt in midjourney), while Nightshade is intended for poisoning training data.
Nightshade should also be significantly less noticeable in the image. Let's hope so at least...
To me, the glazed examples look good enough to use it in addition to the waterwark i use in my online shop.