A few years ago I switched from looking for royalty free images to add to my Daily-Ink blog posts to using AI. The main reason for this is that I found myself spending almost as much time searching for images as I was spending writing my blog post. This was not efficient.
While I’ve had a few challenges along the way, for the most part, I got image creation consistently down around 3-5 minutes. This is great, and so much less stressful… except for when it isn’t. The past few days have been a struggle. I couldn’t get the AI to give me what I wanted. Even when I asked for clarification, and a description of my image was recited ack to me in incredible detail, the end product did not match what I hoped for.
Three of the last four days I was at the gym, walking on the treadmill and I was still trying to get my post published, delayed by failed attempts to get the image I hoped for. In all 3 cases I settled for something close. Actually 2 out of 3, for the other one I used a sample Wikipedia image the AI found for me. I can’t believe how hard it is to get AI to create the image of a clock showing the time 5 o’clock!?!
I was tempted to say that AI image creation is getting dummer, but I think what’s happening is that I’ve just started to expect a lot more. The clock is a bad example, but in many cases I’m expecting a level of sophistication I haven’t asked for previously. I want specific perspectives. I’m asking for complex scenarios, and I’m challenging the AI to create ‘unnatural’ situations, like a teacher in a circle of desks with the students all sitting looking out and away from the teacher.
That’s sounds like an easy request but in millions of reference images of teachers, the AI has been trained to have students face the teacher. So despite continued attempts, with the AI actually describing in detail what I’m asking for before giving it to me, I still got an image of the students facing inward, towards the teacher. Again, and again, and again.
So I’m going to dumb it down. I’m going to ask for less complex images. I’m going to settle for an image that might not be perfect, and most importantly I’m going to spend less time on images and more time writing.
—-
Post script: My one and only image request for this post – ‘Create a stylized, abstract watercolour image that looks like an AI image gone wrong, with an uncanny valley styled mishmash of items.’
