Dove: Standing for real beauty; changing how AI generates images of beauty
Earlier this year Dove took a pledge never to use AI-generated models in its advertising instead of real women.
However, the brand soon realised that by taking this stand, it would still only be able to influence the imagery in its campaigns.
Thousands of brands would still be using cost-effective AI-generated images of women in their advertising, marketing and other media.
And because AI is trained on existing material that promotes stereotypical, predominantly Western-skewed beauty ideals, they would, in the process, be perpetuating the age-old problem.
Dove decided to change that.
Unattainable standards of beauty
Since its inception, mass media has been criticised for promoting unattainable standards of beauty.
From airbrushing to contouring, Photoshop, and more recently, social media filters, the media has long been in the business of promoting a standard of beauty that’s “flawless”, “perfect” and impossible for the average woman to achieve in real life.
Now, with AI-generated imagery becoming more prolific, the perpetuation of perfection as a standard of beauty is being exponentially reinforced.
It’s estimated that by 2025, 90% of the content we engage with could be AI-generated.
Challenging the narrative
Dove’s Real Beauty tool forms part of the brand’s broader goal for everyone to feel beautiful – a goal that has informed their advertising since 2004.
"Twenty years ago, Dove researched how women and girls perceived themselves and found that only 2% thought they were beautiful. That was before filters and social media,” says Lerato Dumisa, marketing manager Dove Masterbrand at Unilever.
“Sadly, two decades later, things are not much better.
“We surveyed thousands of people across 20 countries and found that 1 in 3 felt the need to alter their appearance based on the content they see online, even when they know what they are seeing is not true.
“It has become part of our brand DNA at Dove to challenge the narrative that promotes that kind of thinking.
“Women, and young girls especially, need to see normal people who look like them portrayed as beautiful.
“AI is a powerful tool. We need to use it to do good.”
Reflecting the Nuances of Africa
The African continent – and even just South Africa – is a prime example of how generalisations fall short.
When prompted to generate an image of a South African, AI has a natural skew towards black South Africans.
The Real Beauty tool makes the AI more representative by introducing nuances.
“We have an opportunity to drive further African and South African representation within Generative AI through prompts reflective of the diverse races and ethnicities in our region,” says Dominique Baxewanos, creative director at VML South Africa.
Training representation
Dove and creative agency VML South Africa devised a plan to train AI to include human characteristics typically labelled as flaws, such as age spots, freckles, wrinkles and vitiligo, as well as varying skin tones and diverse ethnic features.
The aim is to teach AI, those using it, and the audiences they’re targeting that real beauty lies in the “imperfect” physical idiosyncrasies that make us human.
To bring the vision to life, software giant Adobe gave the team early access to their Firefly beta API (application programming interface), which they embedded into a closed-system Real Beauty Generation microsite.
Adobe Firefly and Adobe Express are products within the Adobe Creative Cloud suite that designers and artists use to create content via text-to-image AI.
Adobe users generated up to one billion images per month towards the end of 2023.
Through the microsite, users can refine generated images to include more human characteristics using a drop-down menu of prompts.
Users then rate the pre- and post- Real Beauty images.
This rating, as well as the prompt sequence, is sent to Adobe to build on its AI image generation technology.
The team also partnered with body positivity expert Zuraida Jardine-Lindberg, who provided key insights and considerations in refining the prompts.
“For the success of our project, it’s important to strike a balance between diversity in ethnic and racial representation, without bias or stereotype as a baseline to apply our Real Beautymodifiers,” explains Theo Ferreira, executive creative director at VML South Africa.
“We’ve balanced this out with simplified frontend descriptors for the dropdown fields, and more detailed backend descriptors that will be fed to Firefly through the API. The frontend descriptors will be refined with our body confidence expert to ensure we are using the correct language.”
