With recent sky rocking attention and access to consumer-based machine learning, I went to work searching for a meaningful use for this technology. After weeks of exploring and educating myself on the various systems like Midourney, Dall-E2, Stable Diffusion, and others, I felt confident that I was ready to begin an experiment. My first goal was to leverage machine learning to generate a photographic representation of “The crowned head of Nefertiti,” wife of Akhenaton, through the use of self-hosted Stable Diffusion systems based on a single sculptural photo as input. The primary focus being exclusively toward facial restoration.

This was especially interesting to me as it represents a variation of a forensic composite artist. These artists use descriptions of people to generate a likeness. Their success relies heavily on the quality of the description, which must be given in detail to the artist so that they can accurately render the subject. To this end, the forensic artist uses questions and intuitive skill to gain as much insight as possible. Having a precise prompt or description that contains the right combination of information is equally important to generate the proper image when using machine learning applications.

In the case of this project, the source image is a photo by Werner Forman/Universal Images Group/Getty Images. The sculpture was made by the famous sculptor Thutmose and was found in his workshop. Egypt. Ancient Egyptian. Amarna period c 1373 1357 BC. Tell el Amarna. This image serves as a visual prompt for progression of machine generated images.

The entire process required several hundred iterations with specifically selected AI generated outputs used as iterative input to serve as progressive checkpoints to ensure that landmark facial structures didn’t drift away from the original source photo.

I started and finished with the same source image and ran the entire experiment four times from start to finish. Interestingly, this resulted in a facial representation that was nearly identical in all tests. No specific starting seed was used and the same prompt was used, “Nefertiti.”

A selection of 5 progressive transformations have been placed in series to demonstrate the process and final result.