One of the things I love most about AI art generation is how unexpected the results can be. 

Recently, I was experimenting with how an AI generated initial image affects the output alongside the same prompt. 

The idea of using an initial image is that it gives the AI a starting point from which to generate from. I was particularly interested to see if using another AI generated image made a difference to the accuracy of the output because using an ‘organic’ initial image can produce weird or very bias results. 

I opted to use the prompt ‘Last selfie on earth’ because I love anything dystopian (aside from real life ironically now that I’m getting older) and I knew it would produce interesting results based on its popularity.

However, what I didn’t expect was what came next. Otherwise I would have considered not running the prompt at 2am. 

Using NightCafe Studio, I selected two pictures of AI generated women rendered in a digital art style. 

Picture 1 featured a kind of alien anime girl that I nicknamed ‘Galaxy Gazer’: 

While picture 2 featured a digital drawing of a haunted girl with blood on her face which I will refer to as ‘Ellen’ for the rest of this post:

The idea was that the AI would generate outputs featuring female depictions of the ‘last selfie on earth’ but with different ‘feels’, colour schemes and slightly different art styles. But that wasn’t what happened. This is what happened: 

Reader, I don’t think I need to tell you how it feels when it’s 2am and something unexpected like that happens. Especially if you’re a stupid mole person like me who works in darkness with a Spotify playlist that contains more ambient and downtempo music than you’d like to admit. 

After the brief spike in emotion, I was fascinated. I had used two different initial images with the same prompt and somehow, on both occasions, it resulted in an almost identical batch output. 

While there was some variants between the individual outputs, the AI was clear. The last selfie on earth would depict a man with facial hair, olive skin and be no older than his early 40s. 

The output produced from ‘Ellen’ particularly looked like a kaleidoscope of different outcomes for the man in question. In some outputs, he’s young, scared and defensive, as if facing a threat he can see or in the middle of conflict. In others, he’s tired and old, sometimes seeming oblivious or intoxicated, sometimes looking scared and powerless to change his fate from an unseen threat. 

The output from ‘Galaxy Gazer’ was more ambiguous with more variance on the physical features of the man, making him appear to be more like different individuals. However, as mentioned before, both batch outputs are linked by common features like facial hair, olive skin and age range. 

So now, the question remains, why did NightCafe’s stable diffusion AI art generation algorithm do this?

The conspiracy theorists amongst us might say it’s because AI knows something we don’t. It’s thoughts like that which make AI art generation so irresistible to keep picking back up and experimenting with. 

However, the more boring explanation lies with how the AI has been trained to process information and how this affects the output. 

For example, the ‘Ellen’ initial image uses a dark and almost grayscale color scheme and her face is fairly distinctively high angled. When the AI is fed this image, this is the main information that it takes away from it to apply to the output. 

The AI then takes the information and weighs up how to apply it to the prompt.

The AI possesses biases or in media communication terms ‘conventions’ that help it consistently produce work that the user expects to see. 

In the case of the prompt ‘Last selfie on earth’ , the conventions that AI has decided to apply is that the subject should be male, have facial hair (there are never razors during the apocalypse) and be olive/tan skinned (There also isn’t any shade or sunscreen). 

Reader, I could write an entire book about the ethics of conventions, why they exist and if they should even exist but it’s a well trodden niche that plenty of other great content creators are currently exploring. My job as a content creator is to simply point out what the conventions are and why they are present in the image in front of you. 

So in this instance, the AI has judged the subject to be exclusively male because they are almost exclusively the hero in the action and adventure genres. 

This is particularly true for the dystopian sub-genre where the protagonist is a male trying to overcome isolation and a hostile environment that doesn’t cut him a break. So he must become stronger and tougher to survive, and maybe recapture his essential humanity along the way. 

As mentioned before, the man’s appearance is supposed to be shaped by said hostile environment. He’s been wandering a wasteland with the sun beating down on him so he’s got a strong tan. There are no razors or grooming kits because personal grooming is a very low priority for a human living through the apocalypse. So he can’t cut his hair or his facial hair very often or very well. 

The resulting output is a mashup of the main details of ‘Ellen’ and the conventions that the AI holds about the prompt ‘Last selfie on earth’. This is how you end up with a batch output of 9 very similar looking men (because Ellen’s face is distinctively high angled) with a dark color scheme (again because of Ellen) and a disheveled and generally morose appearance (conventions). 

In summary, take caution when using AI image generation at 2am, even if the company who produces it is appropriately called NightCafe studio. You might get a shock you didn’t expect. 

However, if you do get an output that shocks you, don’t panic, AI probably doesn’t know something you don’t. Most outputs are just a heady mix of the initial image you provide and the conventions that AI expects you to see based on your prompt.