It doesn’t retain pixels at all. A better way to describe what it retains (though not accurate) is brush strokes. It retains much more of the information than the raw byte count could imply. It’s effectively compressed by capturing relationships of pixels rather than the pixels themselves.
You and I know that - critics might not. They talk like image files go into the model and stay there. One guy insisted that training was no different from encoding a JPEG, because… numbers.
The information gleaned from any single image is hilariously minimal. It’s insulting that anything recognizable comes out. A prolific artist’s contribution means their style is distinguished from the rest of humanity by only a handful of bits.
It doesn’t retain pixels at all. A better way to describe what it retains (though not accurate) is brush strokes. It retains much more of the information than the raw byte count could imply. It’s effectively compressed by capturing relationships of pixels rather than the pixels themselves.
You and I know that - critics might not. They talk like image files go into the model and stay there. One guy insisted that training was no different from encoding a JPEG, because… numbers.
The information gleaned from any single image is hilariously minimal. It’s insulting that anything recognizable comes out. A prolific artist’s contribution means their style is distinguished from the rest of humanity by only a handful of bits.