/ic/ - Artwork/Critique

>>6507127
Oh look it's this post again.
>Look for highly duplicated images in the data set
>tried to prompt that image 500 times each.
>Of the 350k duplicate images generated 500 times, only 109 actually came back sort of similar.
>109 out of 175,000,000.
>0.0000622857%
Wow. Also the newest model removed a lot of the duplicates. Case is dead in the water.

>>6506726
They took the the top most duplicated photos in the dataset. And still despite generating 175 million images, only 109 "near copies" of training set pictures were able to be recreated, with what they call an attack. Extremely clickbaity abstract. This is exactly why stability deduped the dataset in their 2.0 version so this doesn't happen. Besides, being able to create infringing content does not mean the whole tool and process is infringing. Courts will decide whether using public images constitutes as fair use. I believe it is.

>>6505914
They took the the top most duplicated photos in the dataset. And still despite generating 175 million images, only 109 "near copies" of training set pictures were able to be recreated, with what they call an attack. Extremely clickbaity abstract. This is exactly why stability deduped the dataset in their 2.0 version so this doesn't happen. Besides, being able to create infringing content does not mean the whole tool and process is infringing. Courts will decide whether using public images constitutes as fair use. I believe it is.