I see a lot of people playing around with Microsoft Bing’s new Image Creator, which uses Dall-E 3 as its engine. I’m toying around with AI because, like other people, I find it interesting and fun but also because I’m obliged to familiarize it for work reasons I won’t discuss on the blog.
In any case, I decided to try to get Bing Image Creator to reproduce some scenes from my novel Rags and Muffin. Here we go.
First, I wanted Rags sipping tea in her iconic wingback chair with Muffin curled up nearby. These images, I note, are quite similar in lighting and composition no matter how I tweak them. Also, Bing cannot understand what I mean by a furry, dog-like dragon no matter how I phrase it. I get either a dog or a dragon, no in-between.
Some of the images are more convincing than others, but never once has it given me a picture that didn’t have obvious telltale signs of AI generation.
I tried to get it to generate me a picture of Lady Jeanne, and this is where I consistently ran into trouble. I did get some pictures, but I often ran into “unsafe content” warnings. Naturally, Microsoft doesn’t want anyone using this generator for nefarious purposes. Also naturally, the system is inconsistent about detecting nefarious requests. I never once asked it for anything racy or particularly violent but got hit with warnings sometimes anyway, seemingly at random. Asking for pictures of young girls undoubtedly does that.
I did get several images out of it, and I unfortunately forgot to save most of them while I was playing around, but I do have this at least:
That’s close to how I picture the Lady. I was shocked that the generator allowed me to ask for a gun, especially with a young character. I got several images by making the request repeatedly, with slight tweaks, and I got basically that same face every time. It’s an anime-esque face, which suggests that it’s probably trained on a lot of anime-style DeviantART posts.
However, I made a similar request several hours later and got something more photorealistic and with a lot more character, though the gun is obviously screwed up:
Some playing around suggests it’s the sailor fuku that is prompting the AI to give an anime-style face.
I tried to get some action scenes, and the results are underwhelming. I got dinged repeatedly for “dangerous content,” and at one point, it even gave me a more serious warning indicating that I’d lose access to the service if I didn’t knock it off–and mind you, I wasn’t asking for gore.
I tried to get Rags leaping from a building, but every result lacks energy.
Curiously, it consistently wanted her to be barefoot in these pictures, though it kept screwing up her feet:
I decided to ask for shoes and guns, though that didn’t help much. Interestingly, the shoes more than the gun seem set off the automated content warning. It allows “tennis shoes” but always dings me when I ask for “sneakers.”
Rags running through the street was even less successful:
Finally, I tried something more elaborate, a scene in which Rags is engaged in hand-to-hand combat with a marjara. The results were usually hilarious, but the best one is definitely this:
I frankly love this picture. It looks as if it could be a still from a zero-budget adaptation of the book. Yes, it has some weird artifacts like the feathers on the girl’s sock and the severed fingers on her shoe, but aside from that, it looks remarkably good. The background is awesome. But the best part is that the marjara looks like just some dude wearing pieces of a tiger Halloween costume.
Most attempts at this scene were considerably sillier:
I like this because it’s a sweet little girl randomly kicking a stuffed tiger toy. Also, this might be the best dress that the generator produced during all these attempts.
