The Last Waltz A story of an elephant with a broom

Stranger than fiction

Recently I posted to FB a series of images that were cropped from an AI render and were intended to contribute to diffusing some of the myths that have grown around the supposed potential for AI to act as a significant disruptor in the art-world. What I had hoped to draw the attention of art educators to, was that because of the way digital images are decoded / encoded, that regardless of what AI software you use; within certain pictorial spaces, the limitations of neural networks to accurately render all the permutations of the human figure would become blatantly obvious.

In the course of many conversations someone said “ ……and make sure you know how to prompt effectively with your text commands……”

Somehow, someone had mistaken the cropped examples as full renders and because the represented images were clearly problematic, thought; and probably quite innocently, that if I took this advice that my AI renders would be more successful and I would have a more positive outlook in relation to AI image generation.

I’ll be the first to admit txt to image generation has some good things going for it; not least it’s ability to awaken someone’s imagination to the potentials inherent in the seemingly endless novelty of generative possibilities. However on the other side are the weaknesses inherent in the generation and navigation of machine learning databases. Pitted against that is the genius of the algorithms that pull data aligned to the coding of the txt prompts that lure the operator into believing that somehow the neural network has co-operated in realising their request.

The examples provided in this web document serve the purpose of demonstrating that;

  1. Images referenced are neither stolen or used but encoded in accordance with primary visual characteristics i,e colour, shape and line and that by either defining or blurring those characteristics different results accrue.
  2. Using the machine learning database gives different results as opposed to providing a reference image to work with.
  3. You can get significantly different results with the same txt prompts and reference image.

All images in this document with the exception of ‘3’ were rendered with the same txt prompts.

1. Reference image
This image was created for the purpose of this demonstration. There are some rough figures and non-descript background information. There are some broad tonal areas but the predominant element is line.

2. Again same txt prompts, different reference image.

3. Modified txt prompts, different reference image.

Charcoal drawing circa 1987

In all these image grids it’s easy to see that the resultant renders bear only a passing resemblance to the reference image.

The following are random selections from some 3-4,000 image files written to storage on this particular iPad that are representative of explorations undertaken in developing the prompts I now use as the starting point of new investigations.

In my own way this is what I’m also trying to avoid.