> but we don't need to store the output to recreate it; we need only store the formula or code. This doesn't exempt it from information theory. If a 3GiB model

> but we don't need to store the output to recreate it; we need only store the formula or code. This doesn't exempt it from information theory. If a 3GiB model was trained on 5 billion images, it fundamentally cannot have more than ~5 bits of information per training image on average, regardless of whether those are instructions for its recreation or a compressed copy of it. Over-representation of some images (at the cost of the under-representation of others) is an issue, but it certainly cannot "produce a copy of any of its Training Images—which could number in the billions" as is claimed in the lawsuit against Stable Diffusion. > So the "only 3GB" argument does not mean that sufficient instructions have not been stored via ingesting an artist's works to effectively infringe on that artist's works Copyright applies to fixed works in a tangible medium of expression - and explicitly not formulae, procedures, or instructions like recipes. I don't think there's a bright line here, but this wouldn't be the direction I'd argue in if I wanted to classify machine learning as infringement.