Testing GPT Image-1.5

I tested out GPT Image-1.5 for making picture book illustrations on Leonardo AI. This is the original image that I created on Midjourney:

In the story, the hair grows and grows until it is all over the floor of his classroom and starts swamping the other students. I could not get Midjourney to give me any usable images like this. Here are the ones that I could get from GPT Image-1.5 with some limited testing.

My thoughts:

GPT Image-1.5 did not do a perfect job of adhering to the original image, but did do a reasonable job. It was able to stay very consistent once it was using a reference image that it had created without losing the details. This is a big improvement, but I haven’t tried it over many generations to see how long that holds true.

It is able to make images that are more fantastical much better than other models. While many have gotten very good at making very realistic images, making a marching band made of anthropomorphic instruments, for example, has been out of their reach. This model seems to be significantly better at that kind of image generation (like I said, I couldn’t get any usable images in this vein from Midjourney).

Like many models, it has trouble making interesting facial expressions (all the kids in the second image have the exact same expression). With additional prompting, you could probably get this better.

The cost of the model was more than others on Leonardo, but not exceptionally so. I would bet that it would have a cheaper overall cost when you factor in how many times you have to tweak and generate an image with other, cheaper models in order to get what you want.

I was surprised to see that there seemed to be some weirdness with the hands in the second image (the kid in the back), which I hadn’t seen for a while in any models (maybe it was just a perspective problem).