Adventures in AI
I’ve been working on a project for MS Digital Solutions, creating a series of industry specific templates to start targeting a couple of very specific industries for website development projects.
But I needed to dummy up some product images because I didn’t want to use trademarked products in the sample layouts. So I turned to AI.
I started out with the same prompt in ChatGPT, Google Gemini, and Anthropic’s Claude. It turns out that Claude just isn’t really built for that kind of thing. No problem Both Chat GPT and Gemini were more than happy to make a go of it.
It took 3 tries to get a style I liked for a single product out of ChatGPT. But Gemini had one I liked right away. But you will have that kind of outcome with any designer. So I don’t really hold that against either system.
Where the two systems really started to separate themselves was when we started to make different products but keep the same design elements and just change names, icons, colors and basic things like that. ChatGPT really couldn’t help itself. It just couldn’t rein in the creative juices and was attempting to completely redesign each product.
Gemini, once it figure out the elements, was much faster and consistent in creating the 6 different products in two variations each. It started to wander a bit from what it had built once we started doing combination graphics. I had to remind it a number of times to go back to the saved designs.
What I found, is that if might be a bit of a memory creep. The more I asked it to remember all of the changes, the more likely it was to creep and start making things up.
ChatGPT kept having the problem. I never got a complete and clean set of products. With Gemini, I did eventually get the individual products mostly sorted out, but I did start to see little deviations.
So my solution was to go back, download all of the correct products with the matched elements and then reupload them and make corrections 1-at-a-time on the others. Fixing font deviations and missing lines in the logos and some color deviations.
Uploading correct items for reference seemed to take away some of the “guesswork” and kept Gemini from wandering too far afield from the agreed upon designs. It doesn’t solve all of the problems. But it did solve some.
A couple of observations:
- Both Gemni and ChatGPT are very apologetic when you call them out on continuing to make errors that were corrected earlier. So they seem to be aware that they are deviating. But do not seem to be able to stop doing so.
- If you change up the scale of what you ask for in a single prompt, while you may feel like there are more steps to completion, there are fewer error correction steps so it is much quicker.
- Neither ChatGPT nor Gemini really like to create actual transparencies. They create a fake grey and white checkerboard background that we have seen image editors use to display transparency. But not actual transparencies. But when I asked either to create the product image with a white background neither really had any problems.
- Creating a new chat once I had sold product photos, to make the composite images of each product really helped. Especially in ChatGPT, when I took the images from Gemini there for adding to specific packaging.
Looks like AI doesn’t like to do repetitive tasks just as much as we do. If we don’t stay on top of them, their minds wander and they get a little creative with the instructions they are given.