How Emojer leverages ML to create custom emojis for everyone

The exploding popularity of emojis is a curious topic. The natural progression from text-based emoticons of the early 2000s—flashback to : ) or ; ) —to the 😃 and 😉 of today has turned into a full-fledged language- and geography-spanning cultural meme. A smiley face is a smiley face, no matter where you live. Hence the debut of “The Emoji Movie” in 2016, Apple’s big marketing push for its Animoji tech, and more.

But why stop at just these predefined emojis? If you spend any time on Twitter, you’re bound to see various complaints of why certain people and objects aren’t represented in the official emoji register.

This is just the problem that the team at Emojer, a recent graduate of the Y Combinator Summer 2018 batch, is looking to solve. Originally from Romania, the four-person startup has, in an effort to democratize the emoji, developed an app that enables users to create emojis out of real pictures. The idea was born out of the founding team’s previous business, a successful online animation platform dubbed Marionette Studio. While working on that idea, they found there were a lot of non-professionals using it as an easy way to animate pictures. Based on customer feedback, they took that feature and spun it out into its own product—Emojer.

Creating a custom emoji may sound somewhat simple, but it actually presents tough technical challenges. For example, the app must be able to quickly identify the various features of a body or separate the main image from its background. One of the tactics the team at Emojer has historically used to solve these issues is machine learning. They moved their ML workloads over to Amazon SageMaker, which has helped them test and iterate more quickly.

“Prior to SageMaker, we were just using custom machines, buying our own GPUs for computing to run and train the model,” says Raluca Apostol, Co-Founder and CPO of Emojer. “While it worked, it was expensive and hard to set up. SageMaker helped us to quickly build our deep learning models and made our process cheaper and easier, enabling us to run multiple tests at the same time.”