computer vision - Google Inceptionism: obtain images by class -

in famous google inceptionism article, http://googleresearch.blogspot.jp/2015/06/inceptionism-going-deeper-into-neural.html show images obtained each class, such banana or ant. want same other datasets.

the article describe how obtained, feel explanation insufficient.

there's related code https://github.com/google/deepdream/blob/master/dream.ipynb

but produce random dreamy image, rather specifying class , learn looks in network, shown in article above.

could give more concrete overview, or code/tutorial on how generate images specific class? (preferably assuming caffe framework)

i think this code starting point reproduce images google team published. procedure looks clear:

1. start pure noise image , class (say "cat")
2. perform forward pass , backpropagate error wrt imposed class label
3. update initial image gradient computed @ data layer

there tricks involved, can found in original paper.

it seems main difference google folks tried more "realistic" image:

by itself, doesn’t work well, if impose prior constraint image should have similar statistics natural images, such neighboring pixels needing correlated.