Oh wow. this one is cool. A regular slow dive into the standard image net, slowly ramping up the final iteration count as we go, with an added twist to keep it fresh. We also move up and down through the layers, alternating between 4 and 5 endpoints. trippy as anything, i think i might run this one for 20x as long and show it at a party.
A straight dive into a golden curved iterations by octave 5b layer.
The ducks, geese, fish and thimbles seen are consistent with other dives using the same algorithm, with alternative start images. What this shows is that the octave based scaling within the deepdream def: is responsible for the clustering of output images. And its pretty.
Standing on the edge of madness, we leap, fearless into the chaos.
But Wait! someone left the rotation parameter turned up to 11! did they forget that numpy works in radians, not degrees?
Down the rabbit hole!!!