rokket opened this issue on Feb 06, 2023 ยท 35 posts
WendyLuvsCatz posted Thu, 14 September 2023 at 7:30 AM
regardless of how the training works, one only needs to look at sample sizes to realize nobody's work was stolen.
it scraped 512x512 thumbnails of captioned search results on the web to learn how text related to the pixels and did so agnostically.
It didn't think, this image was painted by Joe Blow or this a 3D render by Wendyluvscatz but rather here are these pixels and it has been described as being a render of a black cat by wendyluvscatz and compares it to say another set of pixels described as a photo of my black cat Sooty by Sassy Witch on Facebook, countless other images with Black Cat in description, post, caption etc and learns what is generally accepted as being a Black Cat.
The only reason one would get similar images to real artists is because they are over represented in search engines with a vast number of results for the same search terms using the exact title etc and you can get those exact same images using Google search, you are literally asking it to render an image by that artist, not a picture of this thing or that.
The amount of training it did on anyone's image is so minuscule unless they are massively famous you would probably run out of 0.000n decimal places trying to fit the percentage on a readable chart.
It would use a term like insert Famous Artist name the same way it would any adjective in any context such as "I really like said artist so drew a scribble that sucked because I cannot draw like said artist etc and this caption on their social media feed gets linked with their stick figure of a Witcher character etc, most named artists do not actually result in images that look like stuff by said artist.
Everything fits into a 4GB checkpoint file downloadable on Hugging Face.
That's some serious compression for all the artworks on the internet
so no your 3D render or digital drawing has not been stolen, it was just looked at, compared to other random images with the same keywords in it's description by you or others and if enough images had some key elements the same, the machine learning algorithm determined these pixels are generally considered to be a cat, a painting, a photograph etc. or these things described as photographs all contain the words Elon Musk so this arrangement of facial features ( a concept itself that was trained) is what is expected if Elon Musk is typed.
the sheer amount of images on the internet and used in the LIAON5b database suggests most of it would be in fact news headline thumbnails, adverts, adult content, products for sale, game screenshots etc, actual art itself would be a small percentage, AI is very good at human faces, goods and clothing, Foods come out well as does architecture, scenery.
AI is like homeopathic medicine, a vial of something was opened in the same room as the water in the sold product.