Some thoughts on AI image generation

15 May 2024 at 9:06 am

Article image for Some thoughts on AI image generation

About a year ago (in 2022/2023), I got myself a Midjourney subscription to play around with AI image generation. Initially it was great fun exploring, but I cancelled my subscription after realizing that it would not be able to make anything resembling what I actually needed. It could make great looking pictures, but nothing specific enough that I could actually use it. Now, a year later, I'm setting up a new website and thought I'd try the builtin AI tools of Elementor.

tl:dr Between Midjourney and the builtin Elementor tools, this seems like a step backwards? You can see the results at the top of this article, but the image generation is less successful then it was a year ago. 

What I'm trying to achieve

For the website, I need an illustration showing how to connect my Radsense product to a hobby motor. I wrote the following prompt:

Create a connection diagram with three elements: a hobby DC motor, a screw terminal with three outputs and a Battery. The Red wire from the battery goes to the leftmost pin of the screw terminal. The middle pin on the screw terminal has an orange wire going to one side of the DC motor. The other side of the DC motor has a black wire going to the negative pole of the battery. The battery should be of the typical rectangular 9V type. 

As far as I can see, this is a precise description of what I want? I had hoped to get something like this as the output:

This image was generated by creating and exporting a good starting point from Fritzing as an SVG and then opening this in Illustrator to tweak. The illustration took about 40 minutes to create.

What I got

The four generated images were quite a bit below what I had hoped for. None of them do even remotely resemble what I asked for. No 9V battery, no motor and nothing resembling a screw terminal. In general, neither of the four images are even remotely usable. What disappoints me even more is that all results show clear signs of the images being literally stolen? I obviously do not know exactly where they were taken from, but it's very clear that there is no real intelligence behind how the connections are made and how the images look.

I think creating diagrams like this can be a good future test of how advanced an AI image generation tool is? Showing how to connect items should be a fairly simple task for an AI. I look forward to the day when I can generate illustrations like this using AI rather than Adobe tools. I'll retry this test in the coming years to measure the progress. Beased on what I understand about AI thus far, I'd guess it to stay at this level for a long time and then make a sudden jump to something that is close to perfection an AI arrives that actually understands the problem to be solved.