Hacker News Reader

It's not compression.

Compression implies the input can be reconstructed from the output (lossy or not), in the case of these ml models the input is the training data and the output is the model. You can't reconstruct even a fraction of that training data using the model alone therefore it is not compression even in the most lossy sense.

The model produced though can be an efficient compressor/decompressor, which produces a lossy output image when given a input of prompt and/or image.

All that aside, the whole human/machine thing is a dumb argument. It's humans that are using the tool. The question shouldn't be does a machine have rights to do X, but rather do humans the have right to use and build such tools?

> Compression implies the input can be reconstructed from the output (lossy or not), in the case of these ml models the input is the training data and the output is the model. You can't reconstruct even a fraction of that training data using the model alone therefore it is not compression even in the most lossy sense.

It's already proven that you can reconstruct at least a small fraction of the training set from diffusion models. It's something quite well known, so could we not die on this hill?

[1] https://twitter.com/Eric_Wallace_/status/1620449942090420224 [2] The paper: https://arxiv.org/abs/2301.13188 [3] Relevant HN thread: https://news.ycombinator.com/item?id=34596187

So just to be sure: the list of URLs + metadata that gets used for stable diffusion is several terabytes. Not the images. Just the list of URLs alone (and a bit of other metadata).

Stable diffusion itself is just 6+ GB, and fits comfortably on my USB stick.

That's one heck of a lossy compression algorithm, sir!

(this thread has more discussion on this line of thinking https://news.ycombinator.com/item?id=37879938 )

> So just to be sure: the list of URLs + metadata that gets used for stable diffusion is several terabytes. Not the images. Just the list of URLs alone (and a bit of other metadata). > Stable diffusion itself is just 6+ GB, and fits comfortably on my USB stick.

Thanks for sharing this info which I'm aware of. However, this fact is not as significant as it might sound in terms of whether it's a lossy compression algorithm.

In most lossy compression algorithms, the compression rate is arbitrary. For example, for an algorithm that based on fourier transform, you can choose only take the first sin wave, or the first 1000 ones (a bit oversimplification here).

So yes, SD is small. Quite miracally small, and its size alone implies some important insights on how human see and read artworks. But this fact doesn't change whether I see it as a lossy compression. (In my previous comment I stated human brain stores lossy compressed data too, so you can see I'm using a broad definition of "lossy compression".)

It's not just "lossy compression" though if you can generate images that were never in the source material. I get your point but it's somewhat misleading analogy.

That's a difficult question because the boundaries of similarity/derived works for copyright purposes are determined by judges and juries based on their intuitions. There's no mathematical similarity testing, and trying to formulate such a thing would be challenging.

What's similar enough to a pop music theme, that has a grand total of a few lines of unique music, to be a copyright violation? How many bars have to be copied, and what kinds of minor variances do or don't avoid a violation? If you're inspired by a haiku, and change 5 of 17 syllables, is that still a copyright violation? Who knows.

> That's a difficult question because the boundaries of similarity/derived works for copyright purposes are determined by judges and juries based on their intuitions.

I believe that's why DALL-E bans some keywords related to alive artists. To show they have "no intention to violate copyrights".

And that's why I'm so worry about that we're heading to a future where open, uncensored models are illegal and closed source AI-as-a-service services are legal. It's not fearmongering: right now, you can't use GPL code in your closed source apps, but you can use GPL code on your server running a service that provides the exact same functions. I believe it has already hugely undermined the original intent of GPL (written in an era before SaaS became popular).

Some AI proponents say ML is the biggest invention since steam machines. I don't know if it's true, but if we end up stuck in a situation where open models are illegal while AI-as-a-service is legal, then it's the biggest step toward a dystopia since steam machines.

To the extent that Stable Diffusion models are "lossy compression", the main one is somewhere between 1 and 10 bytes per image depending on whose answer I use for the question "how many images was it trained on?" (I assume the cause is 1.5, 2.0 and SDXL having different answers and the reporters conflating them). The geometric mean of those is ~three bytes, which is only enough for one single RGB pixel per image.

For all the legal issues — and the artistic flaws — I still find it quite remarkable how good it is at such a small size.

> the main one is somewhere between 1 and 10 bytes per image depending on whose answer I use for the question "how many images was it trained

Here is a catch tho. It's just "by average" several bytes. We can't tell if some images practically contribute 0 bit to the final results while some others contribute more.

(I know this "contribute" word is a little non-sense in the context of ML. But existing lossy compression algorithms are not that different in this sense: if you compress a 1M frames produced by a 3D renderer to a .mpeg video, each frame doesn't contribute the same amount of bytes to the final result.)

I'm amused by the idea that under close examination it might turn out that one of the people suing on copyright grounds, made literally zero bits of difference to the model.

Feels very Douglas Adams.

Machines do not have rights. The question is whether an human with a specific machine has a certain right, as opposed to a human with a different machine.

> human with a specific machine

I assume the entire client+server system constitutes the "machine" in this case, correct? So does "human with" refer to the end user (client side) or the sysadmin (server side)? Maybe one is an accomplice? The machine isn't going to infringe without certain prompting by the end user, just as an inkjet printer isn't going to do so.

> Ml model is basically a highly lossy compressed data format This is a pretty incendiary statement for those opposed to generative models, but more important its not a good interpretation because the intent is not to store a compressed format for restoring the same image, nor can it.