Original article was published on Becoming Human: Artificial Intelligence Magazine
As you are reading this, you’re probably locked down under quarantine for several weeks already and you don’t have much to do. It’s likely that you were thinking about doing something creative with your time, like drawing or making music, but you were never really into that and you feel too lazy to learn. If my guess is correct, I have some good news for you. Newest Artificial Intelligence technology is here to help. In this post, I will describe three awesome tools that enable you to channel you imagination and create something beautiful in collaboration with an AI. I will get a bit technical in the descriptions, but don’t worry, you don’t need to understand that to use the tools. They don’t require any programming knowledge or special skills. The only thing you need is an open mind and a dose of courage.
The website artbreeder.com is a cool new project that lets you create beautiful, dream-like art by making mixtures of pictures and ideas. It gives you the possibility to explore and interact in a simple way with a brilliant technology of Generative Adversarial Networks, or GANs.
If you’re not familiar with it, here is a quick explanation. GAN is a technology created for synthesizing images, based on an interaction between two neural networks. The program is given a set of images, for example depicting dogs. One of the networks tries to create a picture that looks as if it comes from that set. The other one, given that image has to judge whether that image is real or generated. They have opposite tasks, and when they try to outsmart each other, they learn really well how to generate and recognize images of various objects. If you train your network on, for example, dogs and pandas, you can use the learned representations to create something that tries to look as a dog and a panda simultaneously, which results in a beautiful pandog, like this one:
Artbreeder is built on top of BigGAN, a huge generative network trained on 150 GB dataset of pictures, with dozens of different classes including animals, plants, buildings and everyday objects. Later they expanded by adding new models based on landscapes, faces, album covers, anime girls and recently, video game characters. Each picture you create has an internal representation in a continuous space of possible pictures. You can freely move around that space, which means that you can create a mix of a cat, a cauliflower and a jellyfish and adjust with sliders how much of each you want. These sliders also take negative values, which means you can see how the opposite of a cat looks like according to the network. The community aspect of the website is very important — you can take existing pictures and change them by adding new components or mixing them together. Every image has its history available, letting us trace back what it was bred from.
The images created by the network look very bizarre and surreal, in a way that makes them fascinating. It creates objects that often resemble something, but nothing you could name and identify, residing in the fuzzy space in-between, open for interpretations. I have a feeling they somehow touch the deep, unconscious parts of my mind, the same way weird, incomprehensible dreams do.
Trending AI Articles:
If now you are interested in trying Artbreeder yourself, check this video with a more detailed introduction: https://www.youtube.com/watch?v=IlrMkHaCosw&feature=emb_title
Deep dream generator
The website deepdreamgenerator.com offers another great and easy to use tool for visual art — style transfer. This technology also relies on deep neural networks and enables us to take two images — one for content and one for style, and create a blend between them. It brilliantly utilizes the structure of convolutional neural networks that are used for image recognition. In the early layers of processing, the network looks for small, simple patterns a few pixels in size — straight lines, curves, circles. Then, in next layers, it gradually builds more complex representation using these low-level ones, up to a level of complex shapes like faces, cars and animals. The style transfer algorithm modifies an image in such a way that the small, low-level representation become more similar to the ones in the style image while keeping the high-level features as intact as possible. This way, the general shapes and structures of the content image are preserved, but the style changes. Here is an example I made:
I hope you get the basic idea. This tool is as simple to use as it gets —you just choose two pictures, optionally tune a few parameters, and let the algorithmic magic do the rest. At the same time, it’s incredibly powerful. With a smart combination you can create high quality, powerful art that is novel and unlike either of the images you start with. I like to use it to create abstract patterns like the ones you can see below. They are made from pictures of simple things around me, like curtains or plants.
The website is completely free to use and has a great community that can serve as a source of inspiration and cool style images.
AI Dungeon 2
If you have ever played an RPG video game, you probably experienced at some point that although the world seems big and filled with opportunities at the beginning, you later discover that the set of actions you can take is actually very limited. That is not the case with the project called AI Dungeon. It is a text-based AI-powered RPG game, where the possibilities are literally infinite. You don’t have to choose your actions out of any set, but you can type anything you want, and let the artificial Dungeon Master decide what happens next. This way, it is not only a game, but a creative endeavor, a story you create in collaboration with an AI.
The program running the game is built on top of GPT-2, a neural network model that was trained on a huge database of text in English to learn the structure and dependencies of human language. When the famous OpenAI lab first created GPT-2, at first they didn’t want to release it to the public because it was supposedly too powerful and could be easily abused for misinformation. You can type in any scenario, even really absurd, and the model will make some sense of it. Even though it does not have any explicit knowledge about the world and the events, it has enough understanding to make a coherent, convincing story including several characters and a chain of events. Sometimes you can see that it doesn’t really understand how the real world works, but it gives an interesting, surreal vibe to the adventure.
The website has a lot of cool features and is being constantly developed. The initial setting of your story can include: fantasy, mystery, apocalyptic, zombies, cyberpunk or a custom one. There is also a community, where you can share and read the generated stories. Another great thing is a dedicated phone app.
I hope you found these interesting, and maybe will use them to fight boredom and look for inspiration. All these websites appeared in the last 2 years and are still fresh, exploratory and underdeveloped. I’m sure that this field will mature and provide us with more and more AI tools that will open new dimensions for both amateur and professional artists.
If you want to see some more art I make, check out my Instagram: https://www.instagram.com/maksql/
Don’t forget to give us your 👏 !
3 Artificial Intelligence tools to enhance your creativity was originally published in Becoming Human: Artificial Intelligence Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.