Whether you want to generate weird and wonderful AI images from text prompts or create a musical composition in partnership with a computer, there’s a lot to explore. These apps are getting better and better with time, and they can give you a good idea as to what AI can do and where it might be headed in the future. It’s the work of software engineer Philip Wang, and it shows off the power of publicly available Generative Adversarial Networks (GANs). These are dual systems where a series of algorithms or neural network produces an output (in this case, an image), another one assesses its quality, and they repeat the process millions of times. In this platform, the first neural network generates a face from a vast library of data. Then, the second neural network judges how realistic the face is based on similarities to pictures of actual people, and it accepts the image only if it passes a certain threshold. While this tool could ensure the world will never be short of generic stock photos again, there are other, more sinister implications to it. Among those, is the potential to create ID documents for people who don’t actually exist. The neural network has been trained on millions of doodles mined from the also highly entertaining Quick, Draw!. Start by picking a category from the drop-down list at the top right of your screen—there is plenty available, from frogs to sandwiches. Magic Sketchpad knows the sorts of shapes and lines that people tend to make when they’re trying to draw simple concepts like a bird, a ship, or a cat, so it can predict what you’ll draw next and finish the doodle for you. The tool can also help artists to augment their work or provide new prompts for creativity. Maybe one day we could see computers doodling as well as humans do. A traditional approach to a project like this would have involved a programmer coding in hundreds or even thousands of responses to specific patterns a user might play. But AI Duet comes up with its own responses based on a huge database of tunes it has trained on. This gives the program the ability to generate melodies that match a user’s input without any specific instructions. This is another example of how AI can work in tandem with artists to produce new creations, whether that’s for movie soundtracks or background music in games. Theoretically, you could rework one riff an endless amount of times. To generate images, Craiyon pulls in information from millions of photos online and their captions. That means it has a vast visual knowledge of everything from celebrities to national landmarks. You can combine two of your favorite fictional characters in a setting of your choosing, or reimagine a famous work of art in a different style— you’ll soon figure out which prompts work best. The results produced by Craiyon are a little bit rough around the edges for now, but it’s not difficult to see how we could eventually use this technology to generate highly realistic images from scratch using only a text prompt. The site was built by creative technologist David Arcus, and it taps into the Google Cloud Vision API, a machine learning system trained to recognize images based on a vast database. So by processing thousands of pictures of dogs, for example, the AI learns to more accurately spot a dog in other photos. Even Stranger Things will try to identify what’s in the picture you’ve submitted and incorporate it into the finished design, usually with broadly accurate results. It’s quite a simple AI tool, but it shows how we can use databases to teach machines to spot new patterns that aren’t in their training materials. The platform is also a great example of how algorithms can apply a particular visual style to photos to create something new. At this stage, AI can’t really finish novels or even news articles, but given enough data and refinement, these are certainly possible uses for it in the future. We might even be able to complete famously unfinished literary works as the original authors would have done. This is another example of how machine learning enables AI to predict a good response to a question or prompt by analyzing patterns in text. Scroll down the page and you’ll see there are four different examples to have a go with: Cats, buildings, shoes, and handbags. Sketch out your drawing in the window on the left, and click Process to see what the AI makes of it. This is another engine based on a GAN, where two neural networks work in tandem to produce realistic results, and even figure out where the edges of objects in images should be. Turning sketches into realistic photos can be hugely useful in all kinds of areas, from building construction to video game design. And the quality of the results is only going to improve over time as these neural networks get smarter and smarter.
