AI and image generation

In researching this week’s blog entries Artificial Intelligence (AI) was a tool.

Harnessing the power of data to create new images is powerful but it currently has limitations. The images on this page were created using AI. By typing prompts into the AI platform new images were created to illustrate the articles on the blog. Below are two examples made using Pixlr.com

The first set of prompts used were “photograph contemporary textiles nature art science music travel research fashion costume traditions” and some of the results are shown below.

There were hundreds and hundreds of images created with these prompts. Scrolling through the list produced more; theoretically the list is endless. But, and it is a big but, all the pictures felt the same. They used the same saturated colours and the same type of drape on the fabrics. They looked impossible – flawless, perfect and beautiful, but impossible. They were also not as startling as they could be, and felt quite safe. However as a starting point for composition perhaps they are useful.

In the second attempt the prompts used were “combine textiles and music with art and craft in a realistic photograph with people”.

Once again the images all felt the same and this time any people depicted were mainly women, and mainly Asian. This makes me think that the AI training datasets associated with the prompt words come from Asia.

The next set of prompts “combine the craft of weaving with music and art in a realistic portrait” resulted in this…

and the last prompts, in an attempt to get a different outcome (combine the craft of weaving with music and art in a realistic portrait of a man) resulted in this.

The experience of using this particular site to create images to illustrate a point was not particularly satisfying. One of the limitations of AI is the outcomes are only as good as the datasets with which the technology is trained. As you can see, the datasets themselves play into stereotypes (only women are interested in textiles and men play guitars), and do not show a great diversity in their subjects, content or output. Over time this may change.