For my midterm project I want to make progress towards my final project, which is a poster design system created with a machine learning algorithm. For the midterm I plan on presenting my method of gathering data in order to train the machine learning program in a particular style.
Background
My goal is to create a functioning design program where the users can input three sections of text, heading 1, heading 2, and heading 3, and have the program generate a poster based on Josef Muller-Brockmann’s grid system design style. I want to use Muller-Brockmann’s style because it has time-tested appeal and a more algorithmic approach to design, as he based his designs on grids with clean fonts and use of size and color to communicate. Many posters in this style use geometric elements, such as lines, rectangles and circles, instead of images, while employing creative placement of text across the canvas. I think it’s possible to break down these posters into data points that would allow me to train an AI on that data so that it can produce posters in this style.
Motivation
I’m very interested in design and I love creating posters ranging from the artistic to posters for parties, events, and various causes. A lot of the design work I do revolves around the same principles, such as layouts following certain rules or grids, certain fonts that go with the different styles I use, and placement of text and images on the canvas. I would love to create an algorithm for generating posters that can create simple, aesthetic designs following design principles by learning which elements go together. I think an algorithm could learn the principles a designer learns and follows, allowing it to create quality designs similar to that of a designer.
References
I’m inspired by Alibaba’s LuBan AI software, which generates banners and ads for Taobao and Tmall. The software was trained on thousands of well-designed ads and can generate 8,000 banners per second, freeing up Alibaba’s designers to work on more complex projects while the menial work of creating ads with similar layouts and designs is left to the AI. The AI was first trained on several different design elements: copy, product shot, background, logo and decorating artifacts. The team then used reinforcement training to teach the AI which types of designs were “good” and which were not. The designs are then evaluated by design agencies on aesthetics as well as success by how many clicks they get. The program was very successful and now many of Alibaba’s ads are created using this software.