We're continuing our look at the AI landscape in relation to eLearning development - taking the hit to see what works and what doesn't. This week it's AI image generators - should you bother?
There's no doubt that good images can transform your online courses, elevating dry text into an engaging learning experience. They can be used to illustrate a point, show a product in situ or bring an idea to life. "A picture paints a thousand words" and it’s certainly easier to show a diagram of the solar system than write out a description of the planets and their position in relation to the sun. Luckily, you’d have no issues sourcing a graphic of the planets on a stock photo website.
But what if you're after something more specific? What if you need a photo of a blonde female wearing your corporate colours, sitting in a park, chatting to a panda? You aren’t going to find that on a stock photo site. As a lifelong Photoshop user, I could knock one up for you but it’s going to take time, effort and money. Is there a way for AI to create the image for you, in a fraction of the time and at a fraction of the cost? Let’s see...
For this test, I used Fotor, which is one of a host of AI image-generation tools out there at the minute. It’s probably the most novice-friendly that we’ve found out of the bunch and doesn’t require you to jump through hoops to access it or start creating AI art.
Like ChatGPT and text, AI image generators are "trained" by looking at billions of images from across the internet and the descriptions and meta-tags linked to them. By doing this, it learns to create art from text. There are moral and ethical implications to this - what images is it looking at, who owns the copyright etc but that's a blog for another time.
Fotor starts you off with a blank text box that invites you to type in a description of the image you want to make. I began with, “Full body photo of a blonde female wearing jeans, trainers and a cyan top featuring the adaptiVLE logo, sitting in a park, chatting to a panda.” A bit wordy but it's what I want.
At first glance, the results didn’t look too bad, barring the noticeable lack of a panda. However, on closer inspection, you can see that their hands are really weird and one of the faces looks like a Picasso. Our logo is also nowhere in sight. As a side note, AI art generators across the board seem to struggle to render human hands correctly.
I refined my prompt. “Full body photo of a blonde female wearing trainers and a cyan top featuring the adaptiVLE logo. Female is hiding her hands, sitting in a park, chatting to a panda.”
Now we’re getting somewhere. The hands are hidden and one of the images has a panda in it. But look at it - it’s terrifying and is trying to hypnotise you. And still no logo. To be fair, it probably has no idea how to interpret our company name, so I’ll leave that out on the next run.
I try again. “Full body photo of a blonde female wearing trainers and a cyan top. Female is hiding her hands, sitting in a park, talking to a friendly, photorealistic panda.”
We are now 2 for 2 on the pandas! And they are definitely friendlier and more realistic. I wish I could say the same for the ladies. I have no idea where the flute came from, although it does seem to be taming the panda in a Pied-Piper sort of way. The other image looks like we’re getting somewhere near until you zoom in on the face of the woman. Has the panda mauled her?
At this point, I gave up on the AI. It had taken me around 15 minutes to come up with the above images and they were nothing like the idea I had in my head. I don’t expect an AI to read my mind but the weird fingers, faces and odd defects were never going to be something I would want to put out there for an audience. There is also no consistency - it renders each woman completely differently, so if you were wanting a series of images, it isn't really helpful. AI "artists" will argue that you can achieve consistency with practice and 1,000 hours to spare but I - and I assume you reading this - don't have the luxury of that time in our busy schedules.
I knew that on stock photo sites, I could find an image of a woman in a park and drop in a panda and our logo in Photoshop and it would look infinitely better, even if it took me slightly longer to create. So I timed myself searching Freepik for suitable images, downloading them and editing in Photoshop - 28 minutes resulted in the composition below. It isn't perfect but it also isn't going to give me nightmares and it's a lot closer to the original brief.
So for me, AI image generation isn’t there yet. If you don’t have Photoshop, your mileage may vary but I would say that searching for free stock photos from Freepik, Pexels, Pixabay or Unsplash is a better use of your time than endlessly typing prompts into an AI generator. Neither will give you exactly what you need, but at least stock photo models have normal hands and don’t look like they’ve been punched by a panda.
Of course, AI technology is moving fast, so maybe I’ll revisit this blog in a year and see where we are.
Why not give Fotor a go yourself and see if you can top the efforts above? We would love to see what you create!
If you're looking for ways to use emerging AI tech in your eLearning, get in touch to see how we can help.