Hacker News Reader

A diffusion model can't make a copy. That's the whole point. The original Picasso isn't in the model weights.

It has learned to make pixels a particular color to mimic that style, but that's it.

If the model didn't learn anything important from Picasso, it wouldn't be in the training data.

This whole argument of "ah but it doesnt really need it" doesn't hold up. If the model didn't need it, it wouldn't have used it in the first place.

Same thing in Artstation. It was of course propitious for AI scientists to find such a lovely database of high quality imagery, and all so helpfully tagged into categories.

All they had to do was take it.

Of course it learned, that's the point of training.

You claimed the model can reproduce an image from that training data. That's false, and what the judge dismissed.

  “none of the Stable Diffusion output images provided in response to a 
  particular Text Prompt is likely to be a close match for any specific image in  
  the training data.”

  “I am not convinced that copyright claims based a derivative theory can 
  survive absent ‘substantial similarity’ type allegations,” the ruling stated.

Whether using copyrighted data to train a model is fair use or not is a different discussion.

> If the model didn't learn anything important from Picasso, it wouldn't be in the training data.

> This whole argument of "ah but it doesnt really need it" doesn't hold up. If the model didn't need it, it wouldn't have used it in the first place.

I haven't seen anyone making this argument. There's a pretty clear difference between learning something from an image and memorizing it.

There also isn't any illegal with memorizing an image and painting a reproduction. What you aren't allowed to do is sell or distribute that reproduction without a license.

I think it makes more sense to restrict what people are allowed do with ML tools than to restrict what ML tools can do.

> Because a Diffusion model can.

This isn't a given, it is something that has yet to be decided in this case. (Edit: if you look at studies that look for examples of SD reproductions, the best examples still similar to what a human trying to reproduce the image from memory would create)

> Your denial that there is a demonstrable difference between human and machine attention

There are demonstrable differences between different intelligent systems. I have yet to see any demonstration that shows that you can't reproduce human attention with machine. (Though we can't do it yet)

> If you don't recognise it, then answer me - Why can't you paint me a Picasso?

I haven't studied painting or Picasso. There are many people who can paint a Picasso as well as, if not better than, any ML model we have today. There are people who you could take to a gallery show who could go home and reproduce both style and individual works at an equivalent level.

> It learns faster and more accurately than any human,

The word, "faster", here is doing a lot of work. Machine learning can be "faster" in that it can happen in parallel and be scaled to take less time. However humans currently also learn "faster" because they require fewer repetitions or examples to learn. As such, the "learning" derived from a human viewing an image is arguably currently larger.

> Yes. You pay for access.

Is this good faith? I already stipulated that the images are publicly accessible. Are you suggesting that somehow artists should be able to block the AI from viewing a properly licensed instance of a copyrighted image? That pretty much results in a ban on general ai.

>Is this good faith? I already stipulated that the images are publicly accessible. Are you suggesting that somehow artists should be able to block the AI from viewing a properly licensed instance of a copyrighted image? That pretty much results in a ban on general ai.

No it, results in a ban on general ai that doesn't compensate rights owners.

So you have a general AI, it sees someone wearing a t-shirt containing a licensed, copyrighted image.

This AI now needs to pay the copyright holder of that image?

How I think it should work is the sa.e for any other intelligent system. Systems can view publicly available images, memorize them, and even reproduce them for certain fair uses. The systems have to pay for a license to right holders for the non fair uses of reproductions.

You're saying it sees someone as if AI is walking around and this all happened by random chance.

No. What happened was AI scientists deliberately built a giant corpus of training data based off unlicensed imagery that was conveniently pre-tagged - Artstation and other sites of the same type. And it was trained to deliberately create images of the exact same type as it was ingesting. It wasn't "randomly learning about the world" and it certainly did not "stumble upon" these images.

The fact that there was a large corpus of artistic imagery already tagged just revealed itself to be too appetizing for AI training, so a few companies did it in secret, without asking anyone for permission, then hoping to make enough money and VC funding that they would defeat any challenge in court.

So yes, those people who made the original images should get paid.

By that same logic, would the manufacturer of my monitor be liable for displaying the unlicensed image at my request? It's a tool. The end user, inputting variables into the model is generating that image.

> You're saying it sees someone as if AI is walking around and this all happened by random chance.

It's called a thought experiment. You are claiming that processing imagery that is publicly available for viewing consitutes an IP violation and I'm taking that to the logical conclusion.

> based off unlicensed imagery

The imagery on Artstation is licensed. Artstation has a license to display those images publicly. If Artstation did not have that license, they would be the infringing party.

> so a few companies did it in secret

I'm not sure what the basis for this assertion is it was not done in secret. Stable Diffusion was trained using https://laion.ai/blog/laion-5b/

This lawsuit isn't against the people who trained the model, but those who distribute it. Yhe Stable Diffusion model was not released by a for-profit company but was released as open source by a university research group which received funding from VC backed companies.

The IP model you are pushing for is a huge expansion of the already problematic copyright system. It will curtain research and training of publicly available models.

It is already easy as can be to copy images with a computer or digital camera. Stable Diffusion doesn't make it easier to reproduce these images and washing an image through a ML model won't protect you from that reproduction being an infringement.

I think it makes far more sense to use our existing restrictions to regulate usage of the output than to make new restrictions on the types of processing that are allowed on publicly available content.

> Can you make me a copy?

If you don't mind it being as bad as the the result of a Stable Diffusion image being passed on to a half-trained robot arm, sure: extra limbs, even more of a David Cronenberg vibe than Picasso at his weirdest, mixing with mis-attributed ideas from other images that I've associated with the same labels…

If you use DALLE3 the results are astonishing. No more issues with hands and legs. Give it a go.

So do you put a sign outside your gallery, "No Savants"?

> Can you make me a copy?

If I am a decent human artist who's looked at many Picasso works and spent time studying how to reproduce his style?

Of course I can.

Artists copy each other's styles all the time, dude. You can literally go to Deviant Art (or wherever), scroll through, and point out examples of style copying, often much more glaring that anything I've seen out of a diffusion model.