A new experiment from Google is looking to help you sketch images faster and more accurately with the help of artificial intelligence (AI).
The software is called Sketch-RNN, and it’s baked into a straightforward web app.
The idea is simple: Select one of the pre-existing objects, start drawing, and the software will try and guess the best way to automatically complete it.
Sketch-RNN’s artificial mind is trained on a neural network fed with thousands of human-drawn doodles, like the ones found in past Google Brain efforts such as AutoDraw and Quick, Draw!.
In a blog post published earlier this year, Google said that the ultimate goal of these AI efforts specific to computer vision is to train machines to identify and recreate objects with an accuracy that mimics human thinking as closely as possible; in this case, the way we draw and connect lines and shapes when trying to sketch an image of a given object.
Sketch-RNN currently has three demos you can try in addition to the standard one: “Multiple Predict,” “Interpolation,” and “Variational Autoencoder.”
“Multiple Predict” works much like the basic demo, but the software will show you multiple possible outcomes at once. For instance, you can begin to draw the body of a mosquito, and Sketch-RNN will show you a few ways it can complete the drawing.
“Interpolation” is a tricky one; the system takes two random images and tries to interpolate them to give you a hybrid result.
Google uses the example of a bicycle and a yoga position, and while the result – like in this case – may not necessarily make sense, it shows that the machine understands how to draw a third object based on the other two.
The last one, “Variational Autoencoder,” actually asks you to draw a complete image of something. When you’re done, it will try to guess your drawing style and give you nine different possible alternatives based on the way you sketched.
Sketch-RNN may not have the same wow factor as, say, AutoDraw, but it’s nonetheless fascinating to see how much (and how fast) computers are getting good at visual recognition.
If you want to toy with Sketch-RNN, head over to Google’s dedicated website here.