Lerma/ combats AI's DEI bias by training model using the likenesses of its own staff

The Stable Diffusion model will be free and open-source.

Lerma/ combats AI's DEI bias by training model using the likenesses of its own staff

More than 20 employees at Lerma/, the Dallas-based ad agency, recently offered their likenesses to train a generative AI model to create more diverse outputs.

For all its impressiveness, generative AI is particularly bad at reflecting diversity. These models, of course, are trained on biased resources—books, articles and websites that often cater to the experiences of majority populations. The result is an omission of minority and marginalized communities from AI-generated content. 

Hoping to correct course, Lerma/ trained generative AI platform Stable Diffusion with a symphony of different faces and bodies that represent its multicultural staff. The agency snapped some 180 photographs of these 22 individuals over the course of nine hours—portrait, profile, full-body—which were then fed into DreamBooth, a software used to fine-tune AI models.

After a few more training sessions, Lerma/ put its new model, dubbed “LERM@NOS,” to work.

Here’s an example of an image of an employee in different styles:

And another across a different set of styles:

The model can yield text-to-image outputs, as well as image-to-image outputs. It’s too early to make conclusions on the effectiveness of the training, but Lerma/ says the model, dubbed “LERM@nNOS,” works by considering a wider definition of people and their appearances. 

“This is our attempt to combat bias,” said Pedro Lerma, chief executive and founder of Lerma/.

Moreover, the agency not only wants to show creatives that AI can be used in inclusive ways but also to help them make such content, which is why Lerma/ is releasing the model as a free and open-source model for all.

Also read: How Lerma/ bolstered relationships and won business

Beginning Thursday, creators can access the model by querying the activation token “lerma-style” within the Stable Diffusion platform. The model’s code will also be available on Hugging Face, a data library for machine-learning software.

“LERM@NOS” builds on previous efforts by Lerma/ to increase diversity, equity and inclusion. In 2021, the agency launched a tool that calculates brands’ multicultural efforts. Lerma told Ad Age that his agency is still using the tool, and is determining how to integrate AI into its functionality.