It would be far more terrifying if the rate of change was less rapid and there was a higher degree of continuity between frames.

Right now it's very noisy, presumably because the frames aren't really interdependent. The video description mentions a blending technique, but its primary purpose seems to be ensuring the underlying video is not completely overwritten by the "dream" generation feeding into itself.

Here's one to top that: https://vimeo.com/132951073

Wow...that is deeply unsettling. Watching it caused me to experience actual physical stress and anxiety.

I'd like to see this for a network trained for more neutral images like places. See for instance the bottom image of [1]. I imagine it would be cool to continuously zoom and iterate through this image in a video.

[1] http://googleresearch.blogspot.ch/2015/06/inceptionism-going...