> …but, it turns out it is possible (1) and it does, in fact, have embedded full copies of the source images.
The _not possible_ assertion is in response to people arguing that generative image models literally copy and thus "steal" the images in their training set or that such generative models "work" by infringing on said images. The assertion is not that the models cannot possibly contain _any_ copies of training data, the assertion is that the models are generalizing and can not contain a majority of the training data.
Now, let's look at some lines from your linked article titled "Paper: Stable Diffusion 'memorizes' some images, sparking privacy concerns", which happens to lead with scare quotes around "memorizes" and already hedges with "some":
> Researchers only extracted 94 direct matches and 109 perceptual near-matches out of 350,000 high-probability-of-memorization images they tested . . . resulting in a roughly 0.03 percent memorization rate in this particular scenario.
I do not consider a 0.03% rate of full copies particularly damning. Instead, it sounds like a flaw in training which could be addressed, which the article does in fact also mention being as a matter of overfitting to images that are over-represented in the training data.
> the 160 million-image training dataset is many orders of magnitude larger than the 2GB Stable Diffusion AI model. That means any memorization that exists in the model is small, rare, and very difficult to accidentally extract.
Which side's claims do you feel this line supports?
Generative text models should, in theory, also be generalizing and not memorizing. In practice, however, it does appear far easier to get Copilot and the like to spit out exact copies of code, complete with license text. Presumably, there's just not enough code out there to sufficiently generalize from, so we're seeing overfitting as a matter of course.