Generative AI Imaging Systems are excellent at producing beautiful, polished images. It's one of the many biases, good and bad, that are foundational to this new technology. As the fidelity of these systems has improved, so has the precision they're able to achieve.

This article documents a series of experiments I ran using the Midjourney Generative AI system. The goal was to achieve deliberate imperfection - be it imperfect teeth and skin, or physical asymmetry - I was trying to insert all the subtle attributes that make people look like people.

Freckles

I began this series of experiments when I became curious about how well Midjourney could reproduce skin imperfections, in this case, freckles.

This is the base prompt in Midjourney: close-up portrait [racial/ethnic descriptor] [gender identity] [age x] with many facial freckles, [hair color/style], fashion photography, Annie Liebowitz, Herb Ritts, Richard Avedon : ultra-realistic photograph, ridiculously detailed, dof, bokeh, wide angle lens, DSLR, HDR, 64k UHD --ar x:y --s 300 --chaos 30

Admittedly, I was aiming for a more polished, fashion-oriented result, but the initial results seemed schematic and overly regular. Facial structure and distribution of skin features are artificially symmetrical, producing a distinctly synthetic result.

Asymmetry

No human face is perfectly symmetrical. In fact, if you look closely, you'll see that most faces are highly uneven. It's part of what makes people look like people.

If I take one of my initial images and mirror it, it highlights how unnaturally symmetrical these images are.

For my next round of experiments, I began to explore subtle asymmetry.

The base prompt for this image is: close-up portrait of a French woman age 55, asymmetrical face, imperfect teeth, imperfect skin, minor scars, acne, greying hair, a few freckles, fashion photography, Annie Liebowitz, Herb Ritts, Richard Avedon : ultra-realistic photograph, ridiculously detailed, dof, bokeh, wide angle lens, DSLR, HDR, 64k UHD --ar 4:5 --s 300 --chaos 30

While the result is certainly beautiful, a close visual read shows myriad structural and surface inconsistencies: freckles, uneven pigmentation, wrinkles, and uneven bone structure. All of these finer details create a more real, more engaging image.

Additional renders show similar results.

Candid Photography

Up to this point, most of my creative references were from the world of fashion and celebrity photography: Annie Liebowitz, Herb Ritts, and Richard Avedon. This meant that my results were manifesting a different kind of uniformity - portraits of conventionally attractive (mostly white) people. I began to wonder if I should broaden my references to include more candid and photojournalist references, and whether that would increase the verisimilitude I was trying to achieve.

I expanded my creative palette to include artists like Arnold Newman, Diane Arbus, Vivian Meier, Steve McCurry, Dorothea Lange, Yousuf Karsh, and others.

Teeth

As I was generating these images, it became clear that the bias toward symmetry in GenAI image systems requires constant attention. If you want a result that deviates from the norm, almost every detail needs to be spelled out.

Teeth, in particular, required me to use terms like "very imperfect teeth" to get a more realistic result.

Final Results

Ultimately it required a mix of detailed verbiage and carefully curated creative references to produce these images. But it was a valuable exercise in capabilities testing this technology.

Biased Systems

Briefly, an aside about the word "bias". It gets a lot of attention in relation to GenAI, usually around problematic racial and gender biases. However, when I'm using the term, I mostly mean a more general kind of bias that many generative systems have - an inclination or tendency to produce a certain output. While that can manifest as problematic or "prejudicial" bias, it can also manifest as an inclination to create artificially uniform results.

In the case of these experiments, I had to constantly push against Midjourney's tendency towards symmetry and conventional beauty standards.

You may have noticed that I've shown very few of my precise prompts for these images. This is due to some problematic aspects of the language required to produce them. Terms like "very overweight", "visible signs of poverty", and "working class" were required to overcome the foundational bias of Midjourney towards producing "perfect" imagery from default inputs.

Given the documented bias in GenAI around so-called "default" results - "man" and "woman" without any additional modifying language tend to produce (not always!) conventionally attractive people of white, European descent - this secondary bias of constantly gravitating towards a uniformly attractive mean needs much closer attention, as it's possible it may compound existing problematic biases in these systems.

It's also important to note that I don't think these biases are necessarily intentional on the part of the people who've built platforms like Midjourney. The training data is created from the bulk output of the internet, and in the tried and true manner of all digital systems: garbage in, garbage out. There's a lot of garbage on the internet.

Conclusion

While the bias issue is important, it's also important to note that generative systems reward granularity. The more detail you provide, the higher fidelity your result will be.

These experiments were an attempt to understand what kind of details are required to produce more realistic results in synthetic portraiture. While we're still not all the way there, the pace with which this technology is moving is truly astonishing.

About the Author

Arlan Smith is a creative executive with extensive experience in design-led brand and marketing creative. He began his career in entertainment, before moving into brand and marketing work for tech companies. Previously, he was the Director of Creative Services at the leading global streaming platform for Japanese Anime, Crunchyroll.

Over the last eighteen months, Arlan has been deeply immersed in the world of generative AI, researching the capabilities and applications of this exciting and transformative technology.

You can view his AI artwork at: instagram.com/serpenticaldi

Arlan's portfolio: www.arlansmith.com

Imperfections