Random AI: Kitschy Child Portraits

After an initial post demonstrating the use of AI to produce images, a number of readers made their own submissions. This one is from February and I am just now getting to it. In certain ways, these images are impressive, but in other ways there is something off about them. I am mentioning the month of these submissions because the technology is purportedly improving all the time.

This submission is from someone calling herself Nikolina. She set up an Instagram account named kitsch_portret_ai to display her efforts. Her interest is in vintage nostalgic-looking images of subjects who are  “glitzy and gorgeous”.

What struck me the most about this first image is not that this girl has no hair on her head, but that the skin mottling seems a little off. It appears the AI is attempting to simulate the texture of skin but does not understand the natural distribution of pores and freckles in real human beings.

The hair on the last two examples is quite convincing and consistent with Hollywood glamour. But these were clearly not constructed from whole cloth but copied from a specific example used to train the algorithm. In the past few months, lawsuits have been filed from writers, photographers and other artists claiming that AI text and image generation can be shown to come from specific sources and constitute copyright violations requiring compensation.

Beautiful 8-year-old girl topless in profile, gold applications, floral patterns, by Gustav Klimt and Alphonse Mucha

This may the longest title in our database. In fact, it’s a set of instructions—cleaned up for clarity—given to an artificial intelligence (AI) which generates images based on a verbal description. This image was submitted by DS from Germany who wanted to begin a debate regarding the ethics and capabilities of emerging AI.

Stable Diffusion AI et al – Beautiful 8-year-old girl topless in profile, gold applications, floral patterns, by Gustav Klimt and Alphonse Mucha

This particular image was created using a free application called Stable Diffusion which, like many in its cohort, create images from text. This is an early stage of developing AI technologies that simulate human pattern recognition and generation by “training” a computer on a multutide of examples on the internet or other massive database. After “seeing” endless examples of various objects and concepts, the computer is finally able to recognize a car or a chicken or a shoe, etc. which can appear in a myriad of variations/styles.

The technology is far from perfect so some quality adjustments need to be made. For instance, DS usually generates hundreds of images on his computer (about two images per minute) and selects the best ones. The algorithm in this case is not error-free, often producing too many fingers or heads appear. By repeating the process, an acceptable form can be found or some reworking may be required using Photoshop.

The more professional versions charge a fee but there are some free versions besides Stable Diffusion such as Midjourney and OpenAI.

There are many ethical issues that come to mind using this technology, some more obvious than others. And it is important we begin to deal with these sooner rather than later. After all, it is not as if we can put the genie back in the bottle. What control do we have over our own image? What about professional models whose images are used? What compansation is due to living artists whose work is used to train the algorithm? What is the potential for fraud in the art collecting community? What freedoms or limits do there need to be with respect to standards of decency?

These are just the tip of the iceberg and the last point forces us to expound on the issue regarding minors portrayed engaging in illegal activity. Already, many companies have put in safeguards to prohibit nudity, but right now it is a simple matter for competent hackers to override this safeguard. Let the discussion begin!