Several days ago, I came across Artbreeder, which uses machine learning to generate art pieces. As their name suggest, it breeds art to generate new pieces. Formerly known as Ganbreeder, it uses StyleGAN and BigGAN models to generate and modify images of faces, landscapes, paints and others.
After creating an account on the Artbreeder website, I was presented with several genre to choose from. As can be seen in the image below, the choices are intuitive such as Portriats, Anime Portraits and Landscapes.
Spending a few hours over several days, I tried most of the genre and I am very impressed. I could see my artwork from my "profile" at the top navigation bar. The image below are several pieces that I had generated. For example, the photo realistic man at the bottom right was my second piece after the more comic looking girl on the top left.
The interface is simple to use. First choose the parent(s). I could use something randomly selected by Artbreeder or I could select from those that are available in the system. Most of the choices looked user generated; some might have been uploaded. In the screenshot below, I chose 3 parent images. I left Chaos as it is but tweaked Asian to 0.3, White to 0.1 and Latino-Hispanic to 0.1. After clicking the blue button with double arrows, Artbreeder generated 3 images. I could save either of them by clicking on the picture. They look human, don't they. I suspect that the developers may be training the AI system using our input.
There are many tweaks to choose from. I could enter the number, use the slider or up/down arrows to adjust the values. Changing the settings will result in different results. Just refreshing or regenerating using the same settings had also generated different portraits.
Creating Anime Portraits
Creating Anime Portraits is somewhat simpler with lesser settings to tweak. I also learnt that pushing the settings to their limits could generate very surprising effects.
This is as easy as Creating Portraits as it is using a similar interface. Because I was aiming for a more natural view, it took a slightly longer time to tweak the setting for the desired results. Saying so, it was still significantly faster (taking a few minutes to configure and mere seconds to generate) compared to painting one myself; which given my skill (lack of) level will not be completed at all. For this piece, I used only one parent image and the results are amazing.
Creating Sci-fi art
This is a very diverse segment with wide range of choices for parents. I was inspired by the Sci-fi city view and used 2 of them as parents. In the process, I had also added, removed and swapped parents to get results that looked more-city like. There are also 2 scales for parents, namely Content (for portraits this would be face) and Style. I believe this sets the weightage for the AI to consider when generating the new image.
Creating Characters & Furry Creatures
I thought this might be a supplementing or complementing tool for giving ideas to the special effects folks in the movies or games industry. They could build the creatures or characters based off these ArtBreeder creations. Sometimes seeing an image can inspire the eventual desired outcome.
While the ability to morph is not new, it is easy to do in Artbreeder. Here I selected a few portraits to morph and after clicking generate, I was told that they would send me the link when it is completed.
One Small Step towards AI being Creative
This is an intial and one of the many steps where AI may become creative. But creativity is subjective. For the above, I had chosen what makes sense to me for the intended purpose. Artbreed generated abstract pieces too. It is definitely fun to see what can be done. It is FREE!!! So go try it out for yourselves.