The idea was explained in a post by Google Allo Expressions Creative Director Jennifer Daniel on the company’s Research blog. The feature combines the power of neural networks with unique artwork to let you build your own sticker pack from a real selfie.
When you take a photo and submit it to the sticker generator, the app will automatically generate an illustrated version of your face. It designs several customised variants equipped with effects and accessories. Once your character has been created, you can use the stickers the next time you’re chatting with friends on Allo.
Google said the most complex part of the system is accurately recognising your face in the photo. Because selfies are taken in a wide range of environments, the context of the photo can be unclear. This can cause an absence of markers usually relied on by AI facial recognition systems.
To circumvent this issue, Google developed a new algorithm that can identify qualitative features of the face. Instead of individually coding how to respond to combinations of lighting colour, background contrast and facial features, the neural network has the freedom to consider multiple options while it’s being run. Various test results have been rated by humans to improve the network’s accuracy.
Once the visual properties have been determined, the neural network generates a sticker by matching features of your face to the artwork it’s supplied with. At this stage, inaccuracies are less of an issue. Instead, they’re desired so the feature can avoid falling into the uncanny valley, the concept that a “true” human image generated by a computer could be repulsive.
Google instead aims to “break the rules of representation,” rather than reproduce reality. The result is a set of light-hearted, brightly-coloured stickers with exaggerated styles and typical social media props. The central face is identifiably similar to your own though, giving you a new way to communicate with your contacts.
“Rather than aim to replicate a person’s appearance exactly, pursuing a lower resolution model, like emojis and stickers, allows the team to explore expressive representation by returning an image that is less about reproducing reality and more about breaking the rules of representation,” Daniel explained.
The chances of generating the same sticker as another user are remarkably slim. According to Google, there are over 563 quadrillion combinations of the different artwork assets created by artist Lamar Abrams. More are on the way too as Google intends to launch a new set of themes that will let you use stickers to convey a wider range of expressions.
The feature is available now in the Allo app for Android devices. It will roll out to iOS users soon. The innovative use of AI gives the struggling chat client a unique feature to differentiate itself from the crowd, drawing on the customisation and sticker options popular with young users in an attempt to generate a new hit capability.