4 ms·

> The thing to remember is AI literally copies the originals. While it mashes them together in a way that makes the lines hard to see that is in fact what it is

> The thing to remember is AI literally copies the originals. While it mashes them together in a way that makes the lines hard to see that is in fact what it is doing. I see this argument a lot, but it cannot be true. The image database behind Stable Diffusion is 5 billion images resized to 512x512 each and compressed down to a roughly 100,000 gigabytes. But, you can run it on a model file that is as small as 2 gigabytes. The originals are not in there. They have not been compressed to 4 bits per image. I've studied and used AI art generators for a long time now and I can tell you from both a technical and experiential POV, the images they create are not mashups.

Guvante 4y ago
None of the evidence you present here supports your conclusion: 20kb per image is like a jpeg worth of data, so certainly not throwing away anything. A tiny model file without logic as to how you got to it isn't interesting. You could just throw away 99.5% (or whatever the ratio is) of the data and call it good. Mishmash data before you do and you might not even have to do that. This style of AI is all about getting right past the human ability to perceive the mishmash and stop there. Given the underlying algorithm is literally about mishmashing I don't know why you are trying to fight that point.
- corysama 4y ago
  0.5 bytes per image. Not 20,000 bytes per image. Mishmashing implies that small clumps of pixels are in the model file somewhere. That's literally not true and has no relationship to how the algorithm actually works. There's a ton of misinformation about how this works being tossed around. It's causing a lot of anger over ideas that simply false. That's what I'm fighting.
  - Guvante 4y ago
    100,000 / 5 = 20,000 off the numbers given here. Don't quote the 2 GB one that one is disingenuous. That is like copying the dictionary heap sorting by word frequency leaving only the most popular words and saying "see I couldn't have copied the dictionary it is too small to have done that". And no the fact that it has "all" the data isn't material here.

Guvante 4y ago

None of the evidence you present here supports your conclusion: 20kb per image is like a jpeg worth of data, so certainly not throwing away anything. A tiny model file without logic as to how you got to it isn't interesting. You could just throw away 99.5% (or whatever the ratio is) of the data and call it good. Mishmash data before you do and you might not even have to do that. This style of AI is all about getting right past the human ability to perceive the mishmash and stop there. Given the underlying algorithm is literally about mishmashing I don't know why you are trying to fight that point.

corysama 4y ago
0.5 bytes per image. Not 20,000 bytes per image. Mishmashing implies that small clumps of pixels are in the model file somewhere. That's literally not true and has no relationship to how the algorithm actually works. There's a ton of misinformation about how this works being tossed around. It's causing a lot of anger over ideas that simply false. That's what I'm fighting.
- Guvante 4y ago
  100,000 / 5 = 20,000 off the numbers given here. Don't quote the 2 GB one that one is disingenuous. That is like copying the dictionary heap sorting by word frequency leaving only the most popular words and saying "see I couldn't have copied the dictionary it is too small to have done that". And no the fact that it has "all" the data isn't material here.