The recent controversy over TikTok’s new video selfie filter, the viral hit Bold Glamour, has finally dragged the media’s attention away from the hallucinating chatbots that have dominated the world of AI for the last several months. Those chatbots (including ChatGPT, Bing Search, Google Bard, and Meta’s LLaMA) all use something known as a Large Language Model, or LLM, to generate their outputs, and they’ve caught the eye due to both their tremendous power, and the remarkable things that can happen when things go wrong behind the scenes.    

‍

KEEPING UP APPEARANCES

Bold Glamour is widely believed to be the product of a different class of AI, one that has little interest in language and is far more concerned with appearances; a Generative Adversarial Model, or GAN. We say “widely believed” because TikTok’s parent company, ByteDance, has refused to provide details of exactly how their software works, most likely because of the ongoing controversies surrounding both Bold Glamour and the TikTok service in general. Bold Glamour is controversial because it generates a live deepfake of a user’s face, with all blemishes and imperfections digitally removed, and even makes alterations to physical facial features and hair. This vision of what you could look like with cosmetic surgery and extensive makeup has been roundly condemned by influencers for reinforcing unattainable beauty standards, and worsening the pressures that are already facing the young women and girls who make up the vast majority of TikTok’s users.

For all the controversy, it’s also a remarkable technical achievement, and one that heralds a new age of consumer AI products appearing within existing apps on your phone. At first glance, Bold Glamour doesn’t appear that different from the more traditional filters that have dominated services like Instagram and Snapchat since 2016. Snapchat’s infamous “dog lens” selfie filter, which adds dog ears, a nose, and a tongue onto your face, works by calculating the location of your facial features and then tracking your head’s movement to keep the cartoon canine additions in place. Hundreds of variations on the filter’s idea exist, but they all share one major flaw; put anything in front of your face, such as your hand, and the effect glitches out. 

‍

‍

AN ADVERSARIAL APPROACH

Bold Glamour doesn’t try to track where your nose is. It uses a GAN to build a new image of your face, and it does it quickly enough to serve up thirty deepfaked frames every second, sufficient to create a smooth video feed. It’s notable that Bold Glamour is only available on reasonably powerful phones; users with older or cheaper hardware are not offered the filter, which further reinforces the theory that Bold Glamour works through a GAN - but how does a GAN work? 

‍

Well, as the word “adversarial” suggests, a GAN really consists of two neural networks. The first is called the Generator, and its job is to generate new content based on its training data. If you give a Generator a bunch of photos of shoes as its training data, and then ask it to create a picture of a shoe, it will do its best to create a new image that resembles the existing shoe pics. 

‍

‍

The Generator’s adversary is the Discriminator; a second neural network whose job is to determine whether the image the Generator creates looks real or fake. A battle of deception then plays out between these two networks, with the Generator trying to trick the Discriminator into thinking its fakes are real. As the process repeats, the Generator’s fakes improve, until eventually the Discriminator is unable to sort the wheat from the chaff. 

‍

HANDS WHERE I CAN’T SEE THEM

This technique can be used for all sorts of applications, and is not limited to images or video, but most well known GANs deal with visual imagery. The popular AI text-to-image creation services, Dall-E and Midjourney, are both GAN-based systems, using a combination of language models and GANs to convert text prompts into incredibly detailed and convincing images. Text-to-video apps are coming; Meta have announced their forthcoming Make-A-Video software, which looks like a powerful tool, but TikTok may have stolen a march on their rivals with Bold Glamour.   

‍

‍

One of the reasons why Bold Glamour is so effective is that it’s highly limited in scope. It is only interested in changing your face, and as such it avoids many of the traps that can trip up more ambitious GANs. One of the most obvious is the problem of hands; the current generation of text-to-image GANs find hands almost impossible to draw, mainly due to their complexity and the wide variation of ways in which a hand can be articulated. Bold Glamour doesn’t care about your hands, and as such it’s free to concentrate on the things it’s good at. While it’s certainly a worrying application of AI technology that seems sure to make an existing problem worse, it’s also a fascinating example of how targeted use of AI can completely revolutionize existing consumer products, for better or for worse.

‍