The Creative GAN_

Making AI generated art more creative with the use of real-time influences.

Tools used



Computational Creativity
Machine Learning


Fay Beening

01 Context


"What if your fridge can make art?" that is the question that triggered this research into AI art generators. AI art generators keep getting better and the interest in these generators is also increasing. But AI art does not go without the discussion of whether AI art can actually be called art.

For my master thesis I looked into this discussion, more specifically into Generative Adversarial Networks (GANs are an often used learning algorithm in art generators). In the end the goal of this thesis is to make the generative process behave more creative by influencing the learning process with real-world influences. A more detailed description of the research and design here.

"How can a GAN become an autonomous artist that generates creative art trough IoT?""

02 Set-up


We investigate how a GAN can become an autonomous artist that generates creative art through IoT. We do this by reviewing the state-of-the-art within computational creativity and evaluating if the GAN is currently creative. From this examination, we derive requirements for a design that we will refer to as the creative GAN. The new design combines a basic GAN with an interactive IoT system to simulate a more creative process. We build a prototype of the design that can produce unique images.


We set up the new GAN system starting with the hardware. The hardware consists of an Arduino UNO that represents an IoT. We use three sensors connected to the Arduino to act as the IoT environment. The values of the sensors link to images of a specific art style from the WikiArt dataset. For each newly measured value the GAN trains on new data.

Arduino as IoT set-up

03 Results


Below is a timelapse of the art that the GAN generates. You can see how the images are a particular color and particular design for short periods of time and ever changing. These changes occur because of the changes in the environment, In this case light, humidity and temperature.

Generative process over time

Final Product

The model uses the images in the working dataset to make unique images in recognisable designs. We can see that the images produced are abstract, though if we were to assign a style to the images, it would be a combination of cubism and graffiti.

Selected art

Back to Top

Copyright © Fay Beening 2023