3 ms·

If I compress an image with JPEG, I won't get the exact same pixels values in the original image after decompressing because the compression is lossy, but I'm s

If I compress an image with JPEG, I won't get the exact same pixels values in the original image after decompressing because the compression is lossy, but I'm sure you'll agree that the original image is "there". How are NN weights different from the DCT coefficients in a JPEG file?

XorNot 4y ago
The NN weights can't recreate anything without input values - specifically a 512x512 grid of random noise, and a transformed textual prompt. So there's almost a kilobyte of missing data, without which nothing is produced. Would a JPEG file with random noise in it be potentially any given original image? Of course not - even though the decompressor is perfectly capable of recreating one given suitable input data. The "Grokking Stable Diffusion" colab posted here a week or two back makes this particularly explicit - any given 512x512 image can be reversed into a latent-space encoding of that image, and reconstructed from it. The NN weights are necessary but not sufficient to do so - ultimately you end up with a 512 byte number mapping to the expression which reconstructs an image. But that includes images which aren't part of the original training set. [1] https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wFCHhGLFooW_pf1?usp=sharing https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wF...
- trention 4y ago
  You can pontificate all you want about what is contained in the model, the fact is a case related to infringement is going to hit an EU court soon enough and any sort of commercialization of those models will be banned. And good fucking riddance to the AI bros, the monkeys of engineering.
  - XorNot 4y ago
    You seem to think it's all figured out, go ahead and try it. This isn't theoretical, everyone seems to think they definitely know how this will go and we sure could use some court clarity. Slam dunk case right? What's stopping you?
    - trention 4y ago
      I can't "try it" because I am not an injured party, me not being an artist that has published any image ever. That being said, the existing "relationship" between google and news publishers in the context of EU court decisions is already convincing enough about the direction this is going to go down (and google's usage of news didn't even have to cost a single person their job, unlike how this will).

XorNot 4y ago

The NN weights can't recreate anything without input values - specifically a 512x512 grid of random noise, and a transformed textual prompt. So there's almost a kilobyte of missing data, without which nothing is produced. Would a JPEG file with random noise in it be potentially any given original image? Of course not - even though the decompressor is perfectly capable of recreating one given suitable input data. The "Grokking Stable Diffusion" colab posted here a week or two back makes this particularly explicit - any given 512x512 image can be reversed into a latent-space encoding of that image, and reconstructed from it. The NN weights are necessary but not sufficient to do so - ultimately you end up with a 512 byte number mapping to the expression which reconstructs an image. But that includes images which aren't part of the original training set. [1] https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wFCHhGLFooW_pf1?usp=sharing https://colab.research.google.com/drive/1dlgggNa5Mz8sEAGU0wF...

trention 4y ago
You can pontificate all you want about what is contained in the model, the fact is a case related to infringement is going to hit an EU court soon enough and any sort of commercialization of those models will be banned. And good fucking riddance to the AI bros, the monkeys of engineering.
- XorNot 4y ago
  You seem to think it's all figured out, go ahead and try it. This isn't theoretical, everyone seems to think they definitely know how this will go and we sure could use some court clarity. Slam dunk case right? What's stopping you?
  - trention 4y ago
    I can't "try it" because I am not an injured party, me not being an artist that has published any image ever. That being said, the existing "relationship" between google and news publishers in the context of EU court decisions is already convincing enough about the direction this is going to go down (and google's usage of news didn't even have to cost a single person their job, unlike how this will).