Meitu is a Chinese tech and selfie-app company which built its net worth of US$4 billion in large part by making selfies look cute. Last November the company launched “Andy,” an AI bot that repaints selfies with a choice of styles and visual effects.
You may find Andy’s dreamy creations too cute for your taste, but numbers don’t lie: at time of writing a total of 166,743,348 users had employed the artbot’s “tech magic” to beautify their selfies and pictures of popular figures like Elon Musk and US President Donald Trump.
Andy runs on DrawNet, a generative network which breaks down the stylistic differences of each image, categorizing elements like composition and paint strokes. Andy was trained in six months on database of portraits with different age, gender, and race profiles. Now, the more it’s used, the more it improves.
Breakthroughs in computer-generated imagery and machine learning from the company’s R&D division, Meitu Imaging Laboratory (MTLab), have enabled Meitu to monetize its massive user database using facial recognition, image identification, AR, and 3D visual effect applications.
Meitu’s newest AI-powered bot is a virtual dermatologist developed in partnership with Meitu’s cosmetics e-commerce platform. From a photo, “MTskin” can diagnose skin conditions based on texture, tone, pores, and dark circles. It then outputs recommendations for skincare products.
The lab collaborates with the Shanghai Dermatology Hospital — an institution that performs 70% to 80% of China’s cosmetics testing. Hospital experts teamed with engineers to label skin types and conditions. “We want our product to be a recommendation tool that is feasible (in the medical sense) and precise,” Meitu CTO Wei Zhang tells Synced.
There are some challenges. Zhang says for example in low-light conditions it’s difficult for the app to separate acne from noise. External light factors can also affect the pigmentation assessment. In addition to hardware limitations, algorithms sometimes struggle in identifying acne, spots, or blackheads. To this end, deep learning is used in conjunction with traditional machine learning algorithms.
At present most members of the MTLab team are tech-savvy males, while most MTskin users are female. “Fortunately, we also have females on the team who can give advice from their point of view,” says Zhang, explaining that a female product manager might for example adjust how a skin condition is displayed in order to avoid upsetting users.
Within three months of its release MTskin had performed 23 million tests. “This is a preliminary version of the product and we have a two to three-year plan for its finalization,” says Zhang.
Journalist: Meghan Han, Wei Zhou | Editor: Michael Sarazen