Alex Vranas @Alex@vran.as

Small correction: Stable Diffusion has 2.3B 256x256 images, and 170M 512x512 images that it is trained on.

But still, that comes out to less than one byte of model data per image in the training set. There's literally no way to argue that a 256x256 image got compressed down to less than a byte.

It is worth noting to those who are in opposition to AI that the "models trained on my work contain my work" argument isn't technically or legally sound. Better arguments need to be made. If we could cram 1B+ 512x512 images into a single 2GB model, it would represent the single greatest breakthrough in data compression technology in human history. It turns out, this isn't a compression breakthrough, it's just proof that the original data isn't in the model.

I agree that some degree of mimicry isn't infringement - fair use includes some degree of derivation - but I really don't think that training and using a model to deliberately imitate another artist's work represents fair use, especially if it can be found to harm the artist.

The line shouldn't be drawn at "is it an exact copy?", it should be drawn at "would it confuse a reasonable person into thinking this was original art by another artist?".

Yesterday's lawsuits against Generative AI companies like Stable Diffusion and Midjourney were mostly dismissed, and I generally agree with that. There is one case still ongoing between Sarah Andersen and StabilityAI that I expect Sarah to lose on the grounds that her accusation isn't sound (her artwork isn't actually stored in the model).

But I am disappointed that they ruled that copying styles isn't infringement unless it is an exact copy of a work. I feel like this isn't a nuanced ruling...

Sad to remove Cohost from my bio, but I do not ever use it anymore. I don't really have anything bad to say about it, it just wasn't my style.

Still love eggbug though.

@stux I really want to learn how they do it. I have tried a lot of experiments with Stable Diffusion, and it does a terrible job with a 360 view. I'm amazed that Skybox manages to get all the shadows pointing the right way!

@apLundell @jalefkowit No worse than most of the new houses here in Seattle. Lots of 3+ storey townhomes with about 300 square feet of space per floor. Something like 20% of the floorplan is stairwell.

I was legitimately searching for housing a year ago, and gave up when the agent asked me to make an offer the same day without an inspection on a townhome that had the kitchen on the third floor.

@apLundell @jalefkowit If they built this thing, I would never want to live in it, but I would feign interest just to get a tour. It wouldn't be the first time I pretended to have enough money to afford real estate just to tour a stupid looking house.

The worst I ever saw was a house with three different second stories that didn't connect to each other, and they each had a separate stairwell for access...

@apLundell @jalefkowit I typed "McMansion" and "McMansion Indoors" into Stable Diffusion, and gave it way too much canvas to work with, and I think you nailed it.

@esvrld@octodon.social There are also some holes in my philosophy as I consider this. There are a few models out there are unabashedly designed to copy a specific style. I have a Pixar-theme model that I have played with that is undoubtedly trained on Pixar's IP.

But I feel substantially less conflicted in infringing upon a $7.4B company, and should be more nuanced about how such training directed at smaller artists' works could be damaging to those artists.

@esvrld@octodon.social I think your proposal that "ai isn't actually capable of copying either the style or the subject matter of artists in the way that is alleged" is a philosophically interesting idea, but as a cautious AI advocate I would argue that, in practical terms, AI is very capable of copying an artist's style to the level of infringement, and would argue that responsible and ethically informed use of generative models is of high importance to anyone exploring this tech.

@esvrld@octodon.social this is interesting input. It definitely would be an issue that certain niche art styles in the corpus could end up zeroing in non-trivially on a specific artist's style. I hadn't fully considered this, and have been assuming that in the absence of an artist's name in the prompt that the art would be more generic. This immediately has me feeling more skeptical of third party models that don't share their datasets and could be weighted in such a way to be infringing by default.

@stux These are really fun! I'll make sure to clone the git repo so I too can own one of these "NFTs".

Just curious, are the backgrounds generated from Skybox AI (Blockade Labs), or sourced from somewhere else?

And, sorry if this is long-winded I just want to get back around to the topic of generative art.

What if you had never seen a cat before (unrealistic, I know). Now imagine that you look at a handful of drawings, paintings, and photographs of cats to learn what one is.

You then draw a picture of a cat, and your only frame of reference is the small selection of other people's works.

Do you owe credit to the artists and photographers who taught you what a cat looks like?