Latent Diffusion Models - The Basic First Alpha Idea

The original process was very very straightforward. Take a stable diffusion generated image And code it using the same encoding scheme look for neighbors in the training set, right? This is this is like the basic first alpha idea that literally the first thing I could think of. We looked at duplicates and data sets tend to weight generations more heavily to be more perceptually similar They are one of the main drivers of like models being able to exactly reproduce images from the training set So we can we take something like lion Are they trying to scrub duplicates out?