Seeking the imperfect…
We are nearing the release of our ‘Simulated Selves’ installation and have been honing the selection of images for our generative AI image gallery. Once again, the important thing is to get the prompt right. In the beginning, achieving the desired outcome was trial and error. We also had to learn the strengths and weaknesses of different generative AI platforms.
As mentioned previously, Adobe Firefly is great for rendering super realistic landscapes and images of people (albeit with an extra limb here and there). It is not so great at creating more ‘human’ renderings (i.e. imperfect images that look created by hand). DALL-E 3 (courtesy of Bing Chat) also struggles with more gestural and artistic renderings with a tendency towards designerly outcomes. While good for more commercial purposes, the crispness can be a bit of an issue when trying to emulate a more human form of expression - like a slightly wonky painting of a house.
On that note. Here is an example of DALL-E’s interpretation of a wonky house:
DALL-E 3|Bing Chat response to ‘create an imperfect slightly wonky painting of a house’.
As you can see it is quite a lovely watercolour illustration interpretation of the prompt. The request was slightly misinterpreted to apply to the house being wonky rather than the painting itself. Furthermore, the house is really only wonky in the sense of a children’s book rendering. Maybe it is about getting the prompt right… but even expanding on the prompt to elicit a more amateurish outcome was misinterpreted and thus unsuccessful in fulfilling our desired brief.
DALL-E 3|Bing response to ‘create an imperfect slightly wonky painting of a house that looks like it was painted by an amateur’.
My final prompt test to get a ‘wonky' painting by using the prompt: ‘create an imperfect, child-like slightly wonky painting of a house’ also did not hit the desired mark. Like the previous examples, the outcomes are quite sophisticated and use a well-balanced composition and considered colour pallette. It seems that generative AI’s geared towards an ‘industry’
Stable Diffusion is a little better at generating more painterly results that are less perfect and more in line with human artistic expression.
Stable Diffusion response to the prompt: 'create an imperfect, child-like slightly wonky painting of a house’ in artstyle-expressionist.
With user demand, generative AI tools may integrate a ‘human error’ scale to better simulate the imperfect.