Who Defines Beauty: Humans or Meitu?
If you have the largest database of human portraits on the planet and each one is manually beautified, can you use machine learning to generalize beauty for the human race?
Who Defines Beauty: Humans or Meitu?
As science seeks to make sense of the world through a digital lens, the abstract notion of beauty is no exception. Founded in 2008, China’s Meitu (“beautiful picture,” in Chinese) is a Hong Kong-listed company with a vision to become “the tech company that best understands beauty itself” .
Meitu’s approach to understanding beauty is to offer free photo-editing and virtual makeup features to its 450 million monthly active users, including options to enlarge eyes, elongate legs, and slim faces . These enhanced photos train Meitu’s machine learning algorithms on how humans view themselves at their most beautiful, which then feed back into expanding Meitu’s capabilities to accurately identify and automate opportunities to beautify a photo. Meitu thus establishes a virtuous cycle of user data, making machine learning core to its product development, organically training itself from data-driven product additions which then increases customer engagement. Meitu currently averages six billion photos a month, with Meitu-perfected faces dubbed “wang hong lian” (“Internet-celebrity face”) and over half of China’s social media selfies upgraded by Meitu .
In total, Meitu has nearly 200 patents and software copyrights on machine learning and augmented reality . Meitu’s R&D division, Meitu Imaging Laboratory (MTLab), is credited with industry breakthroughs in facial recognition, image identification, augmented reality, and machine learning. Once its free apps created a sizable user base and inflow of data, Meitu sought to monetize its machine learning technology with live-streaming and smart hardware.
Meipai has become the #1 video-based social media platform in China, allowing for easy video customization with refined accuracy in identifying backgrounds, faces, lighting, etc. . Meipai has over 140 million monthly active users , and unlike its previous photo apps, Meitu allows its users to monetize their videos with the platform receiving up to 30% of each user’s advertising generated income .
However, Meitu’s biggest cash cow comes from its smartphones (93% of sales) . Claiming to have the largest database of human portraits on the planet, Meitu takes its machine learning technology one step further with its smartphones’ “auto-beautification” — editing in real-time, as you pose for the camera, with smoother and lighter skin, rounder eyes, and whatever else Meitu has been trained to think of as beautiful (see Exhibit 2) .
Meitu has become so efficient with its instant beautification that a user is hardly given time to decide what’s beautiful before Meitu decides for her. As Meitu’s algorithms get smarter, is the user still training Meitu, or is the machine training the individual? The results of machine learning are only as robust and accurate as the data it’s being fed, with the risk of the machine reflecting biases inherent in the data.
As Meitu expands globally, with over one million users in each of the 39 countries it operates in, it faces global challenges of what beauty means for different cultures . Meitu is learning that Norwegians like freckles, but the Chinese hate all blemishes . Meitu has also been accused of racism for whitening people’s skin, and responded with a skin-darkening feature (see Exhibit 3) .
Rather than championing a specific East-Asian interpretation of beauty and reacting only when there is negative press, Meitu should segment its customers culturally and train separate data sets. From there, Meitu could even tailor its editing features per user based on what the app learns about that user’s preferences — for example, the “smoothing skin” feature could keep or remove freckles depending on what it’s learned from the user’s past actions. Ultimately, how a user chooses to beautify themselves should be a personal choice, not a homogenized suggestion from a phone application.
As Meitu steadily builds out its “beauty ecosystem” by expanding into advertising , e-commerce, external partnerships (including brand ambassadors and dermatology hospitals ), and customized cosmetic recommendations , it needs to be conscious of such potential subliminal messaging. To avoid generalizing beauty across cultures, Meitu should consider partnering with local social influencers and recommending products accessible from that region’s local market. By doing so, Meitu can also strengthen its machine learning data by not mixing data that would contaminate and confuse the algorithm, and refine the accuracy of its data.
This year, Meitu announced its long-term goal of becoming a social media platform —the equivalent of Instagram in China (which doesn’t currently exist) . If Meitu’s theory of beauty were to be layered onto Instagram’s significant cultural power, the consequences could be socially destructive, especially for young and impressionable users. Given the possibility, how should Meitu harness machine learning appropriately and what safeguards should the company create for itself as its social influence on Chinese and global beauty standards increases?
 Meitu, “What is Meitu?”, https://corp.meitu.com/en/about/overview/, accessed November 2018.
 Meitu, “Creating and Sharing Beauty”, http://global.meitu.com/en/company, accessed November 2018.
 Jiayang Fan, “China’s Selfie Obsession”, The New Yorker, December 25, 2017, https://www.newyorker.com/magazine/2017/12/18/chinas-selfie-obsession, accessed November 2018.
[Exhibit 1] Daniel Li, “How to create a virtuous cycle of data with your customers”, VentureBeat, August 19, 2018, https://venturebeat.com/2018/08/19/how-to-create-a-virtuous-cycle-of-data-with-your-customers/, accessed November 2018.
 Faye Brookman, “Meitu’s Makeup Plus Launches Counter App to Encourage Virtual Lipstick Experimentation”, Women’s Wear Daily, June 21, 2017, https://wwd.com/beauty-industry-news/beauty-features/meitus-launches-counter-10923569/, accessed November 2018.
 Meitu, “Meipai”, http://global.meitu.com/en/products#Mp, accessed November 2018.
 Jacky Wong, “Meitu: Why China’s Photo Phenomenon Is Out of Focus”, Wall Street Journal, August 23, 2016, https://www.wsj.com/articles/meitu-why-chinas-photo-phenomenon-is-out-of-focus-1471989783, accessed November 2018.
 Fan, “China’s Selfie Obsession”, The New Yorker.
 Sijia Jiang, “Chinese Selfie App Firm Meitu Reduces Loss, Aims for Expansion Abroad”, Reuters, March 24, 2017, https://www.reuters.com/article/us-meitu-results/chinese-selfie-app-firm-meitu-reduces-loss-aims-for-expansion-abroad-idUSKBN16V1WW, accessed November 2018.
 Tiffany Ap,“Meitu Banking New Growth on Social Platform, Off-line ‘Real’ Beauty”, Women’s Wear Daily, August 9, 2018, https://wwd.com/beauty-industry-news/beauty-features/meitu-new-growth-on-social-media-physical-real-beauty-1202772000/, accessed November 2018.
 Meitu, “History”, https://corp.meitu.com/en/about/history/, accessed November 2018.
 Celia Chen, “China’s biggest selfie app Meitu turns its eye to social networking”, South China Morning Post, April 26, 2018, https://www.scmp.com/tech/article/2141309/chinas-biggest-selfie-app-meitu-turns-its-eye-social-networking, accessed November 2018.
 Viola Zhou, “Meitu confident of global success despite bumps in the road”, South China Morning Post, February 6, 2017, https://www.scmp.com/business/article/2068537/meitu-confident-global-success-despite-bumps-road, accessed November 2018.
 Meitu, Inc., “Beauty and Social Media: Meitu Announces Strategic Pathways for the Next Decade”, PR Newswire, Aug 08, 2018, https://www.prnewswire.com/news-releases/beauty-and-social-media-meitu-announces-strategic-pathways-for-the-next-decade-300693971.html, accessed November 2018.
 “After Creating 166 Million Super Cute Selfies, Meitu’s AI Sets its Sights on Skin”, Synced, November 10, 2018, https://syncedreview.com/2018/01/11/after-creating-166-million-super-cute-selfies-meitus-ai-sets-its-sights-on-skin/, accessed November 2018.
 Luo Weiteng, “Meitu fights to lead the pack in AI race”, China Daily, April 11, 2018, https://www.chinadailyhk.com/articles/187/189/131/1523420357496.html, accessed November 2018.
 Ben Kwok, “Meitu’s plan to become China’s Instagram could be risky”, Asia Times, August 22, 2018, http://www.atimes.com/article/meitus-plan-to-become-chinas-instagram-could-be-risky/, accessed November 2018.
Student comments on Who Defines Beauty: Humans or Meitu?
Irene, I completely agree with your take that beautification should be a person choice and not a homogenized suggestion from a phone application. However, I struggle with how Meipai or any company could achieve this goal. The power of machine learning comes from aggregating data and identifying trends—like what Meitu discovered about Norwegians liking freckles. I think machine learning will struggle to identify an individual’s preferences apart from that of the group.
What an interesting article. Every time I visit Taiwan or China, I notice people emphasizing specific aspects of beauty particular to the East Asian culture. With the advent of social media, the standards seem to have converged on something that everyone strives to achieve. When you see famous “wang hong” stars in platforms such as Weibo and similar rankings, the features are all eerily similar – pale skin, narrow faces, large eyes, and “cute” gestures or expressions. We’ve long known that plastic surgery has been a route that many individuals have gone towards in order to align with what society may consider “attractive” – but we can save the philosophical debate on autonomy and authenticity for another day. I firmly believe that platforms like Meitu have a social responsibility to make sure that they do not contribute to an unhealthy standard of what is considered “attractive” on a global basis. Instead of recommended filters, the application should just provide the filters or editing tools that people can customize, without suggestions on what should be the “right” beauty. If platforms like Meitu do not utilize its influence to try and push back from potentially socially destructive behavior, issues of self-perception and self-confidence will continue soaring.
I really enjoyed reading your post – super interesting. I also agree with David regarding the challenges with this model incorporating machine learning. My biggest struggle is that the perception of beauty is dynamic. For example, within the beauty industry, we see a shift from a bold and colorful look to a more natural and minimal one (hence the rise in Glossier). By the time machine learning picks up on one trend, consumers may move on shortly after.
Great article Irene! This is definitely a very relevant topic in today’s world. I agree with you that beauty should be an individual choice, my only concern is that by using machine learning, some generalized concepts of beauty could be imposed on impressionable young women (and men). By having a standardized ‘auto-beautification’, even if it is tailored to specific cultures, it could still be damaging people’s self-esteem that do not relate with those standards. Given those effects, do you think there should be some regulation or something should be done?
This is a very eye-opening article. I used to use the photo-editing app quite often, yet I didn’t know how much they can make use of our data to this level. Nevertheless, the fact that Meitu is going to become China’s Instagram is really making me feel worried. While Meitu offers a very good photo editing function, it implicitly dictates how society view “Beauty” by encouraging people to edit their photos, especially with the automatic feature. In fact, each person has their own beauty that could not be automated. I would encourage Meitu to become beyond just photo-editing, but to stand for something in the society.
Thanks for this very interesting post. There is a fascinating part of the post which discusses whether the individual is still training the machine learning algorithm or if the algorithm is now training individuals of what the societal definition of ‘beauty’ is. I’m curious to think about how the algorithm will respond to changing trends. If society trends towards different attributes of ‘beauty’, then this could confuse the algorithm as to which attributes to suggest in the editing software. Will it preference the older trend which has a high quantity behind it or the newer trend which has recency? As Meitu expands into different countries and exists over a longer time period, it will be interesting to see the machine learning process evolves.
Irene, thank you very much for this article! I wonder whether the machine learning algorithm could have an input of a set of preferences from every user, so that the algorithm adapts. Some people argue that there are some universal beauty standards based on specific ratios in different paths of the body related to the symmetry of a person. Would we be able to discover a “true” universal beauty if we all input our different preferences?…..