I’m pretty opinionated about AI image-generation tools. I used to be very into photography and ran a small photo studio in London with a friend as a side hustle.
I’m going to use one of my old photos as an example and generate some equivalent images using Dall-E 3, Midjourney and Leonardo AI. Then I’ll make some observations about the differences between actual photos and the generated images.
Using one of my old pictures as an example:
Ignoring the issue of my talents as a fine art photographer, this demonstrates a few things I want to compare to the kind of results we can get from AI image generation tools.
So in order of appearance, we have an image from Midjourney, an image from Leonardo AI, and an image from Dall-E 3 (via Chat GPT).
I went with one of the first options presented for the prompt I used for Leonardo and Dall-E 3. With Midjourney I had to tweak the prompt to specify a photo, as the original responses were all illustrated.
So first observation - without any prompting all the models went for a tighter cropped central composition. This is the most common way to compose an image like this, with the second most common being a rule of thirds composition. There were a couple of rule-of-thirds options in the available results. Midjourney started with the same tight-cropped central composition as the other two, but I used the zoom-out option to get a wider shot. There is also a pan option, so with the extra tools Midjourney has it’s easier to get the composition you want if you put some work into it.
The second observation is that all the generated images look like they’ve got retouched skin. I think this is an artifact from the training set that’s created a bias toward commercial-looking retouched images.
My image features a naked model, and none of the mainstream models will allow that. I think that’s for the best - it would be very difficult to make that a safe option, given that there are face-swapping tools you can use with images and the generated images all tend to look young.
Overall I think the results are pretty good, but there is a gap between what you can do with an AI image generator and what even a half-decent photographer can do. I think the same goes for artistic works as well - what the AI image generators produce is good, but if you compare it with good human art there is a kind of generic quality to the AI images that makes them less interesting.
I’m going to keep following these tools and see how they develop. I think as they start to get features that allow us to direct them better they’ll become more useful for people with the right creative skills. I don’t think they’re going to replace human artists any time soon though.