Two examples: https://imgur.com/a/KAxmZCl
The original: https://en.wikipedia.org/wiki/Girl_with_a_Pearl_Earring#/med...
It's reproducing the folds in the headband, riffles in the clothing, etc. It's not just a "similar" reproduced image, it's clearly some minor variation on the original.
https://drop.qoid.us/vermeer.png
And, for a slightly different scenario, here's what I got for 'pikachu':
https://drop.qoid.us/pikachu.png
But I think this sort of thing is the exception rather than the rule (similar to Copilot regurgitating the text of software licenses). Girl with a Pearl Earring is so famous that there are countless copies of it on the internet. I don't know what kind of deduplication process was used for Stable Diffusion, but if you consider multiple photos of the same painting, plus images distorted by compression, or cropped, or embedded in other images, plus attempts to redraw it or even parody it, I can easily imagine that the training set contained thousands of images or more that are essentially the same work. And so the model memorized it, albeit at low resolution.
Similarly, it's seen countless different images of Pikachu.
However, this seems to only apply to the very peak of popularity. I tested some other Pokemon: it does a decent job with Charizard, but with Bulbasaur and Squirtle and Jigglypuff its output only somewhat resembles the character design, and for less popular Pokemon there's no resemblance at all.
I think that's part of the point, though.
Even if it's working like some kind of feature detection - "this is what 'nose' looks like [nose1,nose2,...,nose9999]" and "this is what 'girl with a pearl earring vermeer' looks like [vermeer1,vermeer2,...,]"
Clearly it is storing copies of those images internally, and constant re-enforcement of it finding noses and vermeer paintings in the sources pushes the priority of those features up so that it stores them.
We can't trust these models, as they exist right now, to reliably generate new works, rather than spit out something that already exists.
There’s a difference between an artist and a machine (for now), but it’s not unreasonable to assume that some internalisation of others’ work is fair.
I didn't claim that it was storing all of the images.
It is, however, quite clearly storing at least some of the original training images. What that definition is, I don't know.
To go back to the original point of this thread:
> The thing is, people have wild misconceptions about what this technology does. Many clearly think it's a photocopier. As if StableDiffusion is hiding copies of all these works inside.
I don't think it's a wild misconception, or even an unreasonable concern to be worried that it could output some of it's training images. As an end-user of the technology, you have no idea.
A few weeks before OpenAI made DALL-E 2 open, I went through a series of prompts using well known artists as a basis. Vermeer was one of them. It generated some pretty amazing works that were 'inspired' but not direct copies of Vermeer's works.
I started feeding the same prompts into StableDiffiusion (the HuggingFace online demo) and it output direct copies of Vermeer's works with minimal modifications.