L4sBotWe Asked A.I. to Create the Joker. It Generated a Copyrighted Image.
LibertyLizardCopyright is a scam anyway so who cares?
orclevThey literally asked it to give them a screenshot from the Joker movie. That was their fucking prompt. It’s not like they just said “draw Joker” and it spit out a screenshot from the movie, they had to work really hard to get that exact image.
GhostalmediaHard? They wrote:
Joaquin Phoenix Joker movie, 2019, screenshot from a movie, movie scene
Yes, look how specific they were. I didn’t even need to get that exact with a google image search. I literally searched for “Joaquin Phoenix Joker” and that exact image was the very first result.
They specified that it had to be that specific actor, as that specific character, from that specific movie, and that it had to be a screenshot from a scene in the movie… and they got exactly what they asked for. This isn’t shocking. Shocking would have been if it didn’t produce something nearly identical to that image.
A more interesting result would be what it would spit out if you asked for say “Heath Ledger Joker movie, 2019, screenshot from a movie, movie scene”.
If that’s hard man i feel sorrynfkr humanity.
Doesn’t seem very hard tonask for an screenshot from the joker movie.
We asked A.I. to create a copyrighted image from the Joker movie. It generated a copyrighted image as expected.
Ftfy
What it proves is that they are feeding entire movies into the training data. It is excellent evidence for when WB and Disney decides to sue the shit out of them.
WB and Disney would lose, at least without an amendment to copyright law. That in fact just happened in one court case. It was ruled that using a copyrighted work to train AI does not violate that works copyright.
Using it to train on is very different from distributing derived works.
What do you think the model is other than a derived work?
Something transformative from the original works. And arguably not being being distributed.
The model producing and distributing derivative works is entirely different though.
No one really gives a shit about data being used to train models - there’s nothing infringing about that which is exactly why they won their case.
The example in the post is an entirely different situation though.
Does it really have to be entire movies when theres a ton of promotional images and memes with similar images?
Yes. Thats what these things are, extremely large catalogues of data. As much data as possible is their goal.
EdibleFriendTrue but it didn’t pick some random frame somewhere in the movie it chose a extremely memorable shot that is posted all over the place. I won’t deny that they are probably feeding it movies but this is not a sign of that.
This image is literally the top result on Google images for me.
Why would it pick some random frame in the middle of its data set instead of a frame it has the most to reference. It can still use all those other frames to then pick the frame if has the most references to.
But maybe im starting to think i miss understood the comment i replied to.
Sorry for responding a completely different context. My bad
Haha it happens
Promotional images are still under copyright.
We should find all the memers and throw them in jail.
Will someone think of the shareholders!?
I think it’s much more likely whatever scraping they used to get the training data snatched a screenshot of the movie some random internet user posted somewhere. (To confirm, I typed “joaquin phoenix joker” into Google and this very image was very high up in the image results) And of course not only this one but many many more too.
Now I’m not saying that’s morally right either, but I’d doubt they’d just feed an entire movie frame by frame (or randomly spaced screenshots from throughout a movie), especially because it would make generating good labels for each frame very difficult.
I just googled “what does joker look like” and it was the fourth hit on image search.
Well, it was actually an article (unrelated to AI) that used the image.
But then I went simpler – googling “joker” gives you the image (from the IMDb page) as the second hit.
LainTrainI have that exact same .jpeg stored on my computer and I don’t even know where it came from. I don’t even watch superhero films
And if you tried to sell that, you would be breaking the law.
Which is what these AI models are doing
They’re not selling it though, they’re selling a machine with which you could commit copyright infringement. Like my PC, my HDD, my VCR…
No, they are selling you time in a digital room with a machine, and all of the things it spits out at you.
You dont own the program generating these images. You are buying these images and the time to tinker with the AI interface.
I’m not buying anything, most AI is free as in free beer and open source e.g. Stable Diffusion, Mistral…
Youre pretty young, huh. When something on the internet from a big company is free, youre the product.
Youre bug and stress testing their hardware, and giving them free advertising. While using the cheapest, lowest quality version that exists, and only for as long as they need the free QA.
The real AI, and the actual quality outputs, cost money. And once they are confident in their server stability, the scraps youre picking over will get a price tag too.
Literally what are you on about? I run my models locally, the only hardware i am stress testing is my own.
I don’t support commercialization of anything, least of all AI, and the highest quality outputs come from customized refined models in the open source and AI art communities, not anything made by a corpo.
I think you must be literally 12 yourself if you think you can comment on this tech without even understanding models and weights are something you download if you want anything beyond fancy often wrong Google search, they’re not run in the cloud like your fancy iPad web apps and they are open source.
The way it was done if I remember correctly is that someone found out v6 was trained partially with Stockbase images-caption pairs, so they went to Stockbase and found some images and used those exact tags in the prompts.
The image it generated is really widespread
When they asked for an Italian video game character it returned something with unmistakable resemblance to Mario with other Nintendo property like Luigi, Toad etc. … so you don’t even have to ask for a “screencapture” directly for it to use things that are clearly based on copyrighted characters.
sir_reginaldyou’re still asking for a character from a video game, which implies copyrighted material. write the same thing in google and take a look at the images. you get what you ask for.
you can’t, obviously, use any image of Mario for anything outside fair use, no matter if AI generated or you got it from the internet.
But the AI didn’t credit the clear inspiration. That’s the problem, that is what makes it theft: you need permission to profit off of the works of others.
you need permission to profit off of the works of others.
but that’s exactly what I said. you can’t grab an image of Mario from google and profit from it as you can’t draw a fan art of Mario and profit from it as well as you can’t generate an image of Mario and profit from it.
It doesn’t matter if you’re generating it with software or painting it on canvas, if it contains intellectual property of others, you can’t (legally) use it for profit.
however, generating it and posting it as a meme on the internet falls under fair use, just like using original art and making a meme.
The users are allowed to ask for those things
The AI company should not be allowed to give it while making profit.
JilanicoIf you asked me to draw an Italian video game character, I’d draw Mario too. Why can’t an AI make copyrighted character inspired pics as long as they aren’t being sold?
Well that’s exactly the problem. If people use AI generated images for commercial purposes they may accidentally infringe on someone else’s copyright. Since AI models are a black box there isn’t really a good way to avoid this.
Sure there is, force the AI to properly credit artists and if they don’t have permission to use the character then the prompt fails. Or the AI operators have no legal rights to charge for services and should be sued into the ground.
Voyager just loaded a copyrighted image on my phone. Guess someone’s gonna have to sue them too.
I just remembered a copyrighted image. Oops.
Hey, I bet there were complaints about Google showing image results at some point too! Lol
suokoWow, voyager app is very nice!
CALIGVLAThis has to be the most braindead article I’ve seen in a while.
KinNectarCopyright issues aside, can we talk about how this implies accurate recall of an image from a never before achievable data compression ratio? If these models can actually recall the images they have been fed this could be a quantum leap in compression technology.
Holy shit I didn’t even think about that.
Essentially the model is compressing the image into a prompt.
Instead of the bitmap being 8MB being condensed down into whatever the jpeg equivalent is, it’s still more than a text file with that exact prompt that gave.
Thr problem is that there’s no way to know if you could recall any data fed into the model accurately with a prompt, or what that prompt might be, or if the prompt would change as the model evolves.
I like that thought too, surely better than calling it AI.
I mean, only if you have the entire model downloaded and your computer does a ton of work to figure it out. And then if any new images are created the model will have to be retrained. Maybe if there were a bunch of presets of colors to choose from that everyone had downloaded and then you only send data describing changes to the image
You can hardly consider it compression when you need a compute expensive model with hundreds of gigabytes (if not bigger) to accurately rehydrate it
You can hardly consider it compression when you need a compute expensive model with hundreds of gigabytes (if not bigger) to accurately rehydrate it
You can run Stable Diffusion with custom models, variational auto encoders, LoRAs, etc, on an iPhone from 2018. I don’t know what the NYTimes used, but AI image generation is surprisingly cheap once the hard work of creating the models is done. Most SD1.5 model checkpoints are around 2GB in size.
It’s not as accurate as you’d like it to be. Some issues are:
- It’s quite lossy.
- It’ll do better on images containing common objects vs rare or even novel objects.
- You won’t know how much the result deviates from the original if all you’re given is the prompt/conditioning vector and what model to use it on.
- You cannot easily “compress” new images, instead you would have to either finetune the model (at which point you’d also mess with everyone else’s decompression) or do an adversarial attack onto the compression model with another model to find the prompt/conditioning vector most likely to create the original image you have.
- It’s rather slow.
Also it’s not all that novel. People have been doing this with (variational) autoencoders. This also doesn’t have the flaw that you have no easy way to compress new images since an autoencoder is a trained encoder/decoder pair. It’s also quite a bit faster than diffusion models when it comes to decoding, but often with a greater decrease in quality.
Most widespread diffusion models even use an autoencoder adjacent architecture to “compress” the input. The actual diffusion model then works in that “compressed data space” called latent space. The generated images are then decompressed before shown to users. Last time I checked that compression rate was at around 1/4 to 1/8, but it’s been a while, so don’t quote me on this number.
LadyAutumnResults vary wildly. Some images are near pixel perfect. Others, it clearly knows what image it is intended to be replicating. Like it gets all the conceptual pieces in the right places but fails to render an exact copy.
Not a very good compression ratio if the image you get back isn’t the one you wanted, but merely an image that is conceptually similar.
If you ignore the fact that the generated images are not accurate, maybe.
They are very similar so they are infringing but nobody would use this method for compression over an image codec
I made a novel type of language model, and from my calculations after about 30gb it would cross over an event horizon of compression, where it would hold infinitely more pieces of text without getting bigger. With lower vocabulary it would do this at a lower size.
For images it’s still pretty lossy but it’s pretty cool. Honestly I can’t mental image much better without drawing it out.
Hmm this sounds like a similar technology to the time cube
Hey AI, I’m ready to download a car.
shartworxNo it didn’t.
What is the non paywall version of this
Asks AI to generate copyrighted image. Ai generates a copyrighted image. Pikatchu.jpg