I believe it's both. The produced images do not encode pixel-perfect details necessary for reconstruction. These details are reconstructed from the imperceptibl

I believe it's both. The produced images do not encode pixel-perfect details necessary for reconstruction. These details are reconstructed from the imperceptible differences in the outputs which act as markers _and_ the model's memorization of the content at those locations.