VML South Africa has developed its own AI image generator. Dubbed Asili AI, the new tool has been built and trained to generate authentic images of people on the African continent. The tool was developed due to the agency’s frustration with AI image generator tools that have a Western and American bias in their training and that struggle to produce authentic images of South African people.
Asili AI stems from the Swahili word ‘Asili’, which means ‘origin’ or ‘essence’. This word beautifully captures the core mission of the AI-powered image generator: to create authentic, culturally significant African imagery that is deeply rooted in the continent’s heritage.
Stock Photography Gap
The absence of authentic African imagery in stock photography is not a new problem. ‘As South Africans we have always struggled to get decent stock photography featuring real South African people as opposed to stereotypical or Westernised/Americanised people,’ said Matthew Arnold, Chief Innovation Officer at VML South Africa. ‘This hasn’t really changed with the explosion of AI. While you can get really great quality image generation now, the lack of authentic South African training material continues to be a problem.’
For anyone working in the creative field, lack of appropriate stock photography can be a costly, time-consuming, and limiting problem, as imagery needs to be shot from scratch. This can turn a simple job into a big-budget production – which may prove to be prohibitively expensive. The team at VML decided to address the problem by creating their own stock image solution.
‘We needed to create our own stock photography that we could use in our commercial work that allows for a much more authentic African feel,’ said Arnold. ‘With AI moving so rapidly it made sense to use AI as part of the solution to this problem.’
‘This word Asili beautifully captures the core mission of the AI-powered image generator: to create authentic, culturally significant African imagery that is deeply rooted in the continent’s heritage,’ said Arnold.
Asili is based on Stable Diffusion’s Flux model. A major benefit of Asili is that it was created and trained by the same people who will be using it. An internal team spent around six months meticulously training the tool to get its outputs to the standard required for agency work.
‘We sourced a large number of open-source commercially ready images, curated them to ensure they reflected real South African people and then categorised and labelled them extensively to ensure our model understands every nuance – from hair length to eye colour. It’s a very intensive training process that is continually being improved and refined on a weekly basis,’ said Arnold
The biggest challenges came from scale. ‘Ensuring we had a big enough dataset to train the model on was paramount, to ensure we generated unique and accurate images,’ said Arnold. ‘Additionally, scaling these images to be able to work for a wide range of commercial outputs, from digital ads to billboards, took a lot of trial and error to find the exact upscaling technique that works with our model to ensure crisp details at large sizes.’
Representation In Stock Photography
The result is an AI image generator that can produce high-quality, photorealistic images of authentic-looking African people in authentic-looking settings.
In the rapidly advancing world of AI, VML believes it’s this kind of resourceful use of the tech to solve their own specific problems that will give agencies an edge. ‘Asili is a proprietary tool available to VML’s clients only,’ said Arnold. ‘It’s accessible by VML team members at this stage and forms a significant competitive advantage for our teams.’
Asili will continue to evolve, making more types of imagery possible. ‘We are always refining our model,’ said Arnold. ‘Our main focus now is increasing the number of specific South African locations (backgrounds) for images and starting to cater for other African countries as well to provide a bigger footprint of authentic African imagery.’
Asili was developed by Sphiwe Dube and Matthew Arnold, with input from Jarred Cinman and various different creative and production teams within the VML team.
VML SOUTH AFRICA
https://www.vml.com/south-africa
