Dense captioning algorithm applied to Deepdream infinite zoom.

Originally made in 2016: https://twitter.com/genekogan/status/748931203609219072 https://vimeo.com/173062236

Densecap is an early image multi-captioning system developed at Stanford Vision Lab in 2015-16. This video was the result of an attempt to make it caption synthetic images produced by Deepdream, well outside of its training data.

Deepdream is a technique originally developed at Google by Alex Mordvintsev for synthesizing images which hyper-activate neural networks, i.e. make them "hallucinate." The infinite zooming video technique used here was originally created by Mike Tyka.

At the time, I was mostly interested in emergence, i.e. the unpredictable artifacts of making two generative systems (one that generates images, and one that describes them) play off one another, as well as poking fun of the limitations of early AI text-to-image systems when placed into unfamiliar domains that they weren't originally trained for.

Press: Wired: https://www.wired.com/beyond-the-beyond/2016/09/showtime-gene-kogan-deepdream-densecap/ Fotomuseum Winterhur: https://www.fotomuseum.ch/en/situations-post/densecap-deepdream/

Deepdream: https://blog.research.google/2015/06/inceptionism-going-deeper-into-neural.html Densecap @ Stanford: https://cs.stanford.edu/people/karpathy/densecap