Stable Attribution generates AI art using noisy vectors, the original numerical representation of the photo, and its accompanying caption. Read to know more.
It's now possible to create art using artificial intelligence (AI), and Stable Attribution is designed to indicate attribution to the artist that created the source image. Anton Troynikov, the co-founder of Chroma, got candid about the AI art generator.
How Does Stable Attribution Work?
After the release of Stable Diffusion and the model's increased accessibility to a larger audience, Troynikov told IEEE Spectrum that he began to pay attention to the conversation surrounding AI art. He soon concluded that both sides of the argument were being spoken over.
He was interested in finding out if there was a technical way to ensure that technologists and creatives did not compete with one another. They developed Stable Attribution to address the issue.
The vectors used in latent diffusion models-basically, a distinct numerical representation for each image-encode image and the text that describes them. The vectors are given random values (noise) by the model during training.
After that, a model is trained to transition from a slightly noisier vector to a slightly less noisy vector. In other words, based on the caption text that goes with each image in its training set, the model tries to replicate the original numerical representation of each image.
He added that the idea was, basically, to reproduce images as similar as possible to the original ones. The generated image wants to be similar to the images that most influenced it by having a similar numerical representation because these numerical representations come from these pre-trained models that convert images into vectors and back.
ALSO READ: Will Musicians Lose Jobs If AI Starts Making Music? Here's What the Experts Say
How Does Stable Attribution Determine The Artists And Creators?
Troynikov said they truly want to be able to trace the source photos' creation back to the original creator. However, they have image URLs, which frequently originate from a CDN (content delivery network). These URLs are also available in Stable Diffusion's public training data set. The owners and operators of the CDNs, as well as the owners and operators of the websites where such photographs are hosted, could establish that link.
On the website, there is a small submission form. People can also submit the creator's name on the site as they have a submission form, then they will try to link it back.
When asked whether generative AI would affect artistic creation, Troynikov offered two scenarios.
One is, by being able to do attribution, one can proportionately compensate the contributors to their training set based on their contribution to any given generation.
The second, which is more interesting, is if one has attribution in generative models, it turns them from just a generator into a search engine. They can iteratively find that aesthetic they like and then link back to the things that contribute to the generation of that image.
Check out more news and information on Technology in Science Times.
