Generate Icons with AI
Qiqin Le & Qiao Zhang
AI-Generated Drawings
With the development of technology, artificial intelligence is increasingly used to generate drawings. An interesting example is Google’s AutoDraw, which combines machine learning with drawings from talented artists to help us draw like professionals. The algorithm will try to match the user’s graffiti with the drawings in its database, and any possible matches will be displayed in a list at the top of the virtual canvas. The user can simply click one of his favorite options to replace his amateur doodles.
Another creative example is Sketch-RNN. Once the user starts to draw an object, Sketch-RNN will provide multiple possible ways to continue drawing the object based on where the user left off. It can also mimic the user’s drawings and produce similar graffiti.
AI-Generated Images
AI can generate not only drawings but also impressive images. For example, DeepDream, a computer vision program created by Google, uses CNN to find and enhance patterns in images, thereby “creating a dream-like hallucinogenic appearance in the deliberately over-processed images”.
Neural Style Transfer is another way to create stunning images with AI. NST is a class of algorithms that manipulate digital images or videos and endow them with the appearance or visual style of other images. For example, it can transfer the appearance of famous paintings to photos provided by users.
AI-Generated Digital Art
AI enables many artists to create digital art in an unprecedented way. A Book from the Sky was a digital art created by Gene Kogan. The neural network learned a generative representation of Chinese characters and explored the latent space of Chinese handwriting by rendering smooth interpolations among groups of characters.
Mike Tyka is a digital artist engaged in sculpture, painting, as well as a researcher at Google that helped develop Deep Dream and Inceptionism. The charm of Inceptionism lies in that, with enough iterations, even random noise can be transformed by a neural network to have a certain structure.
AI-Generated Icons
An icon is a designer’s abstraction of a certain thing/image. Traditionally, designers use Adobe Illustrator to design icons with different line types, line widths, colors，shapes, and figure-ground relationships.
We are committed to exploring new abstractions of robot images using AI technology. Based on the existing robot icons, we apply Pixel Recurrent Neural Network to enhance the experience of icon design, thus providing new possibilities for robots and hoping to inspire other designers.
Taking existing resized 1500 icons as training data, our neural network can decide which color the next pixel is the most likely to be. If the upper half of an image is input in the neural network, the lower half can be soon generated by AI.
Based on such a technique, we firstly hand-draw some icon sketches. Then we put (part of) them into the network and get result image as feedback.
Due to the time limitation, there are still plenty of spaces for future improvement. The immediate modification could be adding 3D generation and/or color supporting. Also, a topic that demands further research is to find a better paradigm for human-AI collaborative creating. The current model we applied is strictly based on a row-majored sequence, which is quite a confinement for artists’ creativity.
