Let's talk about generative models. Those new bits of AI that can create all sorts of images and videos in seconds. Since those tools were made by humans and trained on a large amount of public information, they reflect the numerous biases that people hold. They're like a sponge, absorbing the good and the not-so-good stuff from the world around them.
We're not hiding from this fact, and neither should anyone else. Companies diving into open source models face a common challenge – bias and stereotyping are still huge problems for systems like DALL-E 2 and Stable Diffusion.
Now, let's talk specifics
Restyle's generative model works on the basis of Stable Diffusion, and as is typical of this model, its AI algorithm sees people of color whitewashed. We noticed this issue, after which we also received several confirmations from our users.
We do care
We don't want to live in a world where the term "businessman" is exclusively tied to the depiction of a white man in a suit, and where "housekeeper" is synonymous with a Latin American woman. Here in Ukraine, we have long been fighting stereotypes about “inferior” nations and defending the rights of equality. As a Ukrainian company, we believe that everyone – irrespective of their race, gender, age, or identity – should be able to declare themselves loudly.
What we did to reduce whitewashing cases
Over the past few weeks, our ML crew has been working on crafting a solution that ensures the generated output image matches the skin color of the input image.
The newly implemented approach allowed us to assess skin color more accurately and adjust the commands given to our models. As a result, we have enhanced the representation of all beautiful, diverse skin tones.
Since Stable Diffusion is kind of a black box, we cannot guarantee the reproduction of the correct and exact skin tone in 100% of photos. However, whitewashing cases will certainly be significantly reduced compared to before. We've experimented with various approaches and have selected the most advanced combination of measures. You can witness the difference firsthand.
Thanks for coming on this ride with us 🚀