JWF2NET

About JWF2NET (“Electric Sheep on Steroids”)

“JWF2NET” is a dnn (deep neural network) trained on top of the popular “BVLC GoogleNet” with JWildfire-flame-fractals.
Basically, those flame-fractals are a collection of my own flames I created over the last years, together with some computer-generated mutations of those flames, which were manually selected. A total of 20000 images where used, and about 200000 iterations of learning were applied.

The first version of the dnn (“JWFNET”) was a failure and also crashed my computer, so it is finally lost. But, for the 2nd generation I refined the image base and some params, and finally got the results I was aiming for: letting the computer dream of flame-fractals.

So, when thinking of flame-fractals as “Electric Sheep”  (an idea from the inventor of the flame-fractals Scott Draves), we now have some kind of “Electric Sheep on Steroids” 🙂

Here is one of the early results, you can clearly can see lots of fractal flowers inside:

Some other examples:

About the software (“How to run it?”)

The software to run JWF2NET is a combination of different packages (caffe, python, java, JWildfire) and is still in early alpha state. Depending on the resolution it can run on GPU or CPU, so there is a rather fast “playful mode” and a slow mode for generating final images at higher resolution.

Currently it only features a Java-based scripting-interface and I’m not sure if I will ever publish it as it really is not easy to set it up.