Moderator: Edi
What's rather less cool about these new AI art generators is that few if any of the artists or photographers whose reference photos were added to the dataset were consulted beforehand, doubly so because most of them are or soon will be competing with these AIs for paid commissions.Mastr Blastr wrote: ↑2022-12-13 09:58amhttps://hotpot.ai/art-generator
What it does is browse the Internet for reference photographs, then uses them to create a new picture.
True. Don't they also use stills from films and so on that don't have single artists to attribute them to?Zaune wrote: ↑2022-12-17 05:28pmWhat's rather less cool about these new AI art generators is that few if any of the artists or photographers whose reference photos were added to the dataset were consulted beforehand, doubly so because most of them are or soon will be competing with these AIs for paid commissions.Mastr Blastr wrote: ↑2022-12-13 09:58amhttps://hotpot.ai/art-generator
What it does is browse the Internet for reference photographs, then uses them to create a new picture.
Its worse than that. These AI were originally based on image recognition AI, and by design if you train one on a single image it will always give you... a copy of that image, with perhaps a bit of noise due to the random number seed, but still a copy. The diffusion model basically takes the training data and assigns a probability to each individual pixel being a certain shade and hue correlating to specific key words, which in turn are assigned to specific image in the training set. I saw one person knowledgeable in how it works describe it as a "probabilistic photocopier" because in principle, each image in the training data has a non-zero chance of being duplicated perfectly by the AI if the right word combination and random seed happens to be chosen. Most AI researchers charitably refer to this phenomenon as "overtraining", but in truth given their origins as image recognition AI repurposed to generate images, its a feature not a bug. This person didn't have particularly kind words to say about the culture of AI researchers, incidentally.Zaune wrote: ↑2022-12-17 05:28pmWhat's rather less cool about these new AI art generators is that few if any of the artists or photographers whose reference photos were added to the dataset were consulted beforehand, doubly so because most of them are or soon will be competing with these AIs for paid commissions.Mastr Blastr wrote: ↑2022-12-13 09:58amhttps://hotpot.ai/art-generator
What it does is browse the Internet for reference photographs, then uses them to create a new picture.
Honestly the best way to use this tool is if you're creating an 2D anime or cartoon, in which you train the AI the art the team has created, so it can help create animation for the "boring" parts of animation, thus saving time for artists.Formless wrote: ↑2022-12-17 11:28pmIts worse than that. These AI were originally based on image recognition AI, and by design if you train one on a single image it will always give you... a copy of that image, with perhaps a bit of noise due to the random number seed, but still a copy. The diffusion model basically takes the training data and assigns a probability to each individual pixel being a certain shade and hue correlating to specific key words, which in turn are assigned to specific image in the training set. I saw one person knowledgeable in how it works describe it as a "probabilistic photocopier" because in principle, each image in the training data has a non-zero chance of being duplicated perfectly by the AI if the right word combination and random seed happens to be chosen. Most AI researchers charitably refer to this phenomenon as "overtraining", but in truth given their origins as image recognition AI repurposed to generate images, its a feature not a bug. This person didn't have particularly kind words to say about the culture of AI researchers, incidentally.Zaune wrote: ↑2022-12-17 05:28pmWhat's rather less cool about these new AI art generators is that few if any of the artists or photographers whose reference photos were added to the dataset were consulted beforehand, doubly so because most of them are or soon will be competing with these AIs for paid commissions.Mastr Blastr wrote: ↑2022-12-13 09:58amhttps://hotpot.ai/art-generator
What it does is browse the Internet for reference photographs, then uses them to create a new picture.
I've seen what happens when an AI gets trained exclusively on a single artist's work, and the "probabilistic photocopier" analogy holds up remarkably well. While any given image might not look identical to any one work by that artist, its clear that specific parts of the image are being taken from specific images in the training data, making this feel awfully like some form of copyright infringement. In fact, Microsoft, GitHub, and OpenAI are currently in a class action lawsuit over developing a similar AI that copies code instead of images, and the training data contained code covered by several open source licenses which use copyright law to keep that code open source. So depending on how that lawsuit works out, Stable Diffusion could very well also be sued for using training data that they admit simply came from trawling Google Images without consulting anyone about the copyright status of their search results. I've done prompts in Stable Diffusion's demo that pretty conclusively prove that copyrighted imagery is in the training set (try Space Battleship Yamato, for instance, and the AI will try to recreate the faces of characters from that show as well as recreations of the titular spaceship; and I've tried other anime, too, with similar results). Its absolutely disgusting that these are presented to the public as "art" AI when they literally don't have the ability to create, only copy. Whatever the law says about them in the end, they aren't ethical to train this way or to use.
There is already at least one comics convention in New York that has banned imagery generated by AI as being essentially counterfeit merchandise. I fully support this.
Watching Cooridor Crew videos with modern animators, apparently they already use a lot of software like you are talking about, but what they actually use them for is predicative modeling to take a character already in one frame and transform that model between frames, basically cutting the number of frames the animators have to generate the old fashioned way. But it doesn't work like the diffusion model in that its not a digital collage, its much more simple and elegant than that once you realize that even seemingly 2d characters these days are usually 3d models the animators are manipulating and then drawing shading over top of. The diffusion model and others like it are way too complicated for professional work, and their downsides become apparent very quickly. For instance, they can't draw backgrounds for shit. Its one of the easiest ways to spot an AI image, in fact. Good looking foreground subject, bizarre or abstract background or environment. If you are animating in the Japanese style, the environment and backgrounds are often some of the most elaborate and expensive parts of the animation process, which can often be mitigated just by using 3d modeling. The western style of animation already doesn't care so much about that, so other software solutions that already exist serve their needs much better.
Manga artist are using it to speed up their work.Formless wrote: ↑2022-12-18 09:36pmWatching Cooridor Crew videos with modern animators, apparently they already use a lot of software like you are talking about, but what they actually use them for is predicative modeling to take a character already in one frame and transform that model between frames, basically cutting the number of frames the animators have to generate the old fashioned way. But it doesn't work like the diffusion model in that its not a digital collage, its much more simple and elegant than that once you realize that even seemingly 2d characters these days are usually 3d models the animators are manipulating and then drawing shading over top of. The diffusion model and others like it are way too complicated for professional work, and their downsides become apparent very quickly. For instance, they can't draw backgrounds for shit. Its one of the easiest ways to spot an AI image, in fact. Good looking foreground subject, bizarre or abstract background or environment. If you are animating in the Japanese style, the environment and backgrounds are often some of the most elaborate and expensive parts of the animation process, which can often be mitigated just by using 3d modeling. The western style of animation already doesn't care so much about that, so other software solutions that already exist serve their needs much better.
I really can't think of many use cases where these kinds of AI are actually useful. They lack any kind of real creativity because they lack the embodied cognition that a real artist has, they just aren't embedded in the real world so they can't imagine an object not already in their training data. An animator or comic artist needs to be able to invent new images, not just recycle old ones.
