therealjendavis asked:

i was wondering if i can use frank's art for personal use (pfp, for example?) she made this cool miku pride thing with what looks like the bigender flag and i can't find any mention in the frank-faq of how, like. copyright works w her ai.

Sure!

In general, I’m fine with people reproducing Frank content in other places as long as they do it with attribution.

And for something like a profile pic, where people aren’t likely to assume it’s your own work in the first place, I don’t care as much about attribution. It’d be nice if you cited Frank in your bio, though. (I’ve seen “pfp by nostalgebraist-autoresponder” on one or two bios already, actually.)

I should put this in the FAQ… I haven’t updated that thing in a while. Probably too long.

ramshacklefey asked:

Have you heard of Loab? And if so, do you have any theories about what the deal with her/it is?

nostalgebraist:

Someone else also asked about this too …

Here’s the original twitter thread about “Loab,” for those who haven’t seen it.

I originally learned about the phenomenon via this response, which provides a partial explanation that sounds plausible to me. In brief:

• There are reasons to expect that, if you go looking for images that are maximally different from another image (or from a prompt), you will run into an element of a relatively small set.
• If you imagine all possible images as a point cloud in a high-dimensional space, these are the “corners” of the cloud’s space.
• If you start anywhere inside the cloud, and try to go as far as possible from that staring point, you’ll end up at one of the corners.
• Loab was discovered using negative guidance, generating an image as dissimilar as possible from a given prompt. So she may be one of these “corners.”
• The person who found Loab went on to generate more pictures of her by asking the model to combine the Loab image with other images.
• I’m not sure how exactly they did this – there are multiple techniques out there – but they would all involve looking for images that are “close” to the Loab image. Per the above, the Loab image is “far” from almost every image except itself.
• So it makes sense that, if you ask for images similar to the Loab image, you get an unusually high amount of similarity – not just the same topic or style, but the same recognizable face, etc.

In other words, this image is a case where the model is unusually bad at semantic generalization. It usually sees images as related in varied, continuous ways, and can form all kinds of hybrids where one side of the hybrid is less or more apparent. But with a Loab-like image, it can’t find a way to preserve only some of the traits, or preserve them only partially; it thinks every picture is either a picture of Loab, or utterly unlike a picture of Loab.

Such cases are presumably rare. But they’re easy to discover in spite of their rarity, because negative guidance hones in on them.

This doesn’t explain why pictures-of-Loab have the specific traits they do, like being violent / gross / horror-movie-esque. Some possibilities:

1. There’s nothing to explain. The “corner” points have to look like something, after all; this is no more or less surprising than anything. (And perhaps the spookiness helped this example become a meme, so the “explanation” is that if Loab weren’t spooky, you wouldn’t have heard of her.)
2. Perhaps it is somehow the result of some data filtering step that tried to prevent disturbing images, or tried to focus on aesthetically appealing ones? For example, perhaps a data set was filtered to exclude disturbing images at some point, and a Loab-like image “slipped through the cracks,” causing the model to memorize it as a weird special case unrelated to the rest of the training distribution.

The latter hypothesis is complicated by the fact that there are really 2 models here, assuming this is some kind of CLIP-conditioned model, which it probably is. There’s the training data/process for CLIP, and the training data/process for the generator. These likely involved different datasets and may not have been done by the same people.

The Loab discoverer doesn’t say which model they were using. It’s very likely that it was a CLIP-conditioned model, unless it was one of the newer Google ones, which I doubt they’d be allowed to tweet about in this way. But even that is conjecture.

(EDIT: it might be a DALLE-1 / DALLE-mini style model, which would complicate much of the above analysis, unfortunately.)

They do note that Loab was later reproduced (in some form?) in Stable Diffusion, which is CLIP-conditioned, so there’s that.

—-

Now that Stable Diffusion is out there, it would be very easy to test the “corners” hypothesis. Just do negative guidance with a bunch of prompts, and see if you can (1) recover Loab, (2) recover Loab from multiple starting points. It would be especially fun to try to find any non-Loab corners and play with them.

Hmm, I’ll try that out later today probably.

Update: nope, I can’t seem to reproduce Loab with Stable Diffusion, or find any “corner” images.

When I use large negative guidance weights with Stable Diffusion, the results seem similar to unconditional sampling.

Compared to unconditional sampling, the colors are more intense, and seems like there’s a weak preference for certain topics (pictures of lawns and houses, sports games, food?), and the images are crisper. But I’m not getting anything too weird or repetitive.

Unconditional, 100 DDIM steps

“Joe Biden campaign tour,” -20 guidance weight, 100 DDIM steps

ramshacklefey asked:

Have you heard of Loab? And if so, do you have any theories about what the deal with her/it is?

Someone else also asked about this too …

Here’s the original twitter thread about “Loab,” for those who haven’t seen it.

I originally learned about the phenomenon via this response, which provides a partial explanation that sounds plausible to me. In brief:

• There are reasons to expect that, if you go looking for images that are maximally different from another image (or from a prompt), you will run into an element of a relatively small set.
• If you imagine all possible images as a point cloud in a high-dimensional space, these are the “corners” of the cloud’s space.
• If you start anywhere inside the cloud, and try to go as far as possible from that staring point, you’ll end up at one of the corners.
• Loab was discovered using negative guidance, generating an image as dissimilar as possible from a given prompt. So she may be one of these “corners.”
• The person who found Loab went on to generate more pictures of her by asking the model to combine the Loab image with other images.
• I’m not sure how exactly they did this – there are multiple techniques out there – but they would all involve looking for images that are “close” to the Loab image. Per the above, the Loab image is “far” from almost every image except itself.
• So it makes sense that, if you ask for images similar to the Loab image, you get an unusually high amount of similarity – not just the same topic or style, but the same recognizable face, etc.

In other words, this image is a case where the model is unusually bad at semantic generalization. It usually sees images as related in varied, continuous ways, and can form all kinds of hybrids where one side of the hybrid is less or more apparent. But with a Loab-like image, it can’t find a way to preserve only some of the traits, or preserve them only partially; it thinks every picture is either a picture of Loab, or utterly unlike a picture of Loab.

Such cases are presumably rare. But they’re easy to discover in spite of their rarity, because negative guidance hones in on them.

This doesn’t explain why pictures-of-Loab have the specific traits they do, like being violent / gross / horror-movie-esque. Some possibilities:

1. There’s nothing to explain. The “corner” points have to look like something, after all; this is no more or less surprising than anything. (And perhaps the spookiness helped this example become a meme, so the “explanation” is that if Loab weren’t spooky, you wouldn’t have heard of her.)
2. Perhaps it is somehow the result of some data filtering step that tried to prevent disturbing images, or tried to focus on aesthetically appealing ones? For example, perhaps a data set was filtered to exclude disturbing images at some point, and a Loab-like image “slipped through the cracks,” causing the model to memorize it as a weird special case unrelated to the rest of the training distribution.

The latter hypothesis is complicated by the fact that there are really 2 models here, assuming this is some kind of CLIP-conditioned model, which it probably is. There’s the training data/process for CLIP, and the training data/process for the generator. These likely involved different datasets and may not have been done by the same people.

The Loab discoverer doesn’t say which model they were using. It’s very likely that it was a CLIP-conditioned model, unless it was one of the newer Google ones, which I doubt they’d be allowed to tweet about in this way. But even that is conjecture.

(EDIT: it might be a DALLE-1 / DALLE-mini style model, which would complicate much of the above analysis, unfortunately.)

They do note that Loab was later reproduced (in some form?) in Stable Diffusion, which is CLIP-conditioned, so there’s that.

—-

Now that Stable Diffusion is out there, it would be very easy to test the “corners” hypothesis. Just do negative guidance with a bunch of prompts, and see if you can (1) recover Loab, (2) recover Loab from multiple starting points. It would be especially fun to try to find any non-Loab corners and play with them.

Hmm, I’ll try that out later today probably.