A stylized robot face made of the letters F and T.

The Devil's Candy


Snoot Entertainment asked us to create footage for the 2015 horror film The Devil's Candy using Deep Dream technology to overlay hallucinogenic images on selected video footage. We used a technique where we provided a satanic-looking "guide image" to produce sinister goat heads rather than the "puppy slugs" that Deep Dream tends to create by default. The footage was ultimately not used in the movie.


To get the unique goat effect, we used a then-brand-new technique for "guided Deep Dream" imagery. Instead of using the typical L2 optimization, this method attempts to match a target image on each generative step. We ran this per-frame on the video footage on an Amazon EC2 GPU cluster. It took almost two days to render all the footage!

To learn more about how to accomplish the guided Deep Dream, take a look at the "Controlling Dreams" section of this ipython Notebook.

Post Mortem

As people who like movies and don’t make movies, it was super cool to work with people who make movies! Maybe we’re supposed to act all chill about it? But we aren’t chill. It was so cool! Likewise, it was wonderful to have an excuse to rent many, many times more cloud processing power than we ever had before; if we hadn’t, we might still be rendering some of these frames on our desktop. (It tried its best!)

Ideally when doing something a little off the wall like this, it’s good to be as close as possible to the creative team. The next time something like this comes along, we’re going to insist on some screen-share time with the director or production staff so we can walk them through some of the broader strokes of what ranges of effects are possible. Deep Dream can be used in a myriad of ways to make freak-ass lookin’ stuff, and all the back and forth of rendering really ate into our ability to demonstrate the possibilities. In the end, we did end up with this effects reel, though, which... sleep well.


The content of this article is provided under a CC-BY Creative Commons Attribution 4.0 license.