A few weeks ago I decided to replace my profile picture. I use the same picture across almost all the social media sites. Now a part of the new profile picture is also the header image in this blog’s design. Here is this new profile picture:
It looks very stylish on the first glance, but if you look closer you will notice that it contains imperfections. The reason for this is that it was generated by AI without any further editing. While I did some basic design work in the past, I am not an illustrator and would never be able to come up with something like this myself. I do not even know how which tools to use these days.
As an alternative to just using a photo or drawing something myself, I decided to make use of AI to generate something that matches my interests of computers, bikes, dogs, beer, and coffee. The goal was to come up with a profile picture I can identify with. I went for the Bing Image Creator by Microsoft to generate images based purely on a textual description. After many attempts, I came up with the right prompt to achieve the above result:
I still have only a vague idea how these AI image generators work, but the results are really impressive. The most important part is to come up with the right prompt and I guess prompt-engineering will become its own profession in the future. In this case, I started with asking for a “profile picture”, but that will make the AI try to create real faces or avatars. Instead I had to go for a more abstract description of what I wanted. I also tried many image generation styles such as asking for an image in watercolor, charcoal, comic, or even pixelated style.
Also many of the generated images had very obvious errors, for example the spokes of the bike were sometimes distorted and a complete failure. You can still spot such defects in the result above, but here I thought they were not as noticable on the first glance. Finally, there is a lot of luck involved to get an image to your liking, because the results are not reproducible. Even with the exact same prompts, the generated images will always look different to those generated before. I do not know whether if Bing Image Creator adds some randomness to the input to reinforce this effect as it is in their interest to offer many different designs to chose from.
Below is a collection of the alternative generated versions I did not select as profile picture. These show at least a bit of variation in style. In others I tried to include more details like a bicycle helmet or plants in the prompt. But adding more specific objects did not improve the results.
It is not that I need to create a brand out of myself, but the green fish I used up to now became a bit boring. This GIF was originally pixeled by me for a forum over 20 years ago.
The animation and its small size did not work well on many sites, so I mostly used a upscaled PNG version anyway. The only reason I still keep it on my X/Twitter account is that the site stopped accepting animated GIFs many years ago. But existing animated profile pictures were allowed to stay. That means I have one of the very few animated profile pictures on the site overall (although animated GIFs might be coming back?). As a side note, I am still pondering whether to delete my Twitter account, as I will not post there anymore. I have fully switched to Mastodon by now.
As I had created the previous profile picture myself, I am the copyright holder (or with a slightly different notion, I hold the Urheberrecht in Germany). However, for an AI generated image, the question of copyright and ownership is not yet fully settled. The issue is that copyright applies to the intellectual conceptions of the author (in Germany § 2 (2) UrhG “Werke im Sinne dieses Gesetzes sind nur persönliche geistige Schöpfungen.”).
To create this new profile picture, I merely wrote the prompt that was given to a tool as input. Is that actually different from giving mouse and keyboard inputs to a tool such as Adobe Illustrator to create something? However, Bing Image Creator was trained on existing art and image descriptions to learn how to create new artworks, so that involves more than what I contributed to the result. According to their terms of service, I am allowed to use the generated results “for any leimagegal personal, non-commercial purpose”. As I do not earn any money with my online presence, the use as a profile picture is within these limits. These will be an interesting challenge for law makers, but I expect that technology will once again advance faster than the definitions of what is legal. I do not want to go deeper into what this means for artists here, but you will find plenty of discussions on this topic on the internet.
The creation of the new profile picture was an interesting endeavour that filled a few hours of leisure time playing with the image generation of the Bing Image Creator. I can recommend to try it out to bring your imagination to life with words only.