If a model is 64 GB, and it’s been trained on one billion images, the contribution of each image is approximately sixty-four bytes. The model contains as much unique trace of each work as a checksum. Thumbnails weigh one hundred times more. And the more works you shove in, the smaller the fraction gets.
People are upset by this because it says rude things about how distinct any human work might be. But the same way you, a human, could describe an artist’s linework, identify its characteristics, and roughly imitate them… the model can apply that as a vector. Those features are what it relates to that name. There’s not a cache of reference images it can glance at, to copy from. That would be so much easier. Instead, it has a mechanical description of how that artist’s works differ from the general idea of art. From the n-dimensional soup of every JPEG ever posted.
But let’s say we did it how you want.
Say we tossed all this out, and only trained on a tiny set of varied images, created specifically for this training. We create a bespoke from-scratch model that’s free and clear of any concerns for authorship, copyright, permission, or content.
That model could still copy any artist’s style. It’s just a vector. Some guy uses thick lines, purple shadows, extreme perspectives? Then so long as the new model has concepts of line width, shadow tint, and foreshortening, all you gotta do is probe for how it classifies that guy’s work, versus anyone else’s. This is how people can share weights to reproduce characters for some anime that was announced yesterday. If you, a human being, could describe Hatsune Miku in terms of her distinguishing features, the model will satisfy those labels.
Training is transformative use.
If a model is 64 GB, and it’s been trained on one billion images, the contribution of each image is approximately sixty-four bytes. The model contains as much unique trace of each work as a checksum. Thumbnails weigh one hundred times more. And the more works you shove in, the smaller the fraction gets.
People are upset by this because it says rude things about how distinct any human work might be. But the same way you, a human, could describe an artist’s linework, identify its characteristics, and roughly imitate them… the model can apply that as a vector. Those features are what it relates to that name. There’s not a cache of reference images it can glance at, to copy from. That would be so much easier. Instead, it has a mechanical description of how that artist’s works differ from the general idea of art. From the n-dimensional soup of every JPEG ever posted.
But let’s say we did it how you want.
Say we tossed all this out, and only trained on a tiny set of varied images, created specifically for this training. We create a bespoke from-scratch model that’s free and clear of any concerns for authorship, copyright, permission, or content.
That model could still copy any artist’s style. It’s just a vector. Some guy uses thick lines, purple shadows, extreme perspectives? Then so long as the new model has concepts of line width, shadow tint, and foreshortening, all you gotta do is probe for how it classifies that guy’s work, versus anyone else’s. This is how people can share weights to reproduce characters for some anime that was announced yesterday. If you, a human being, could describe Hatsune Miku in terms of her distinguishing features, the model will satisfy those labels.