I think this will be my last post on this subject at least for a while, lest this simply turn into a blog of AI-generated art, but I want to report on what happened when I tried to get Bing to produce images of the protagonist from Jake and the Dynamo.
I did this in two different sessions and got very different results each time, though I notice certain similarities running throughout. This isn’t exactly scientific, so I can’t tell you the exact prompts I used, but each one was a variation or slight modification of this:
Skinny ten-year-old electricity-themed magical girl, powder puff tutu, blue and gold breastplate, blue and gold metal boots, bare midriff, simple gold tiara, long blue pigtails, green eyes, holding shield and spear, motif of heart with lightning bolt through it appearing on shield and armor, riding a flying snowboard over an Art Deco city.
I got hit by the “dangerous content” warning inconsistently and seemingly at random, though it appeared more often if I specified either “photorealistic” or “anime-style.” Undoubtedly, the young age of the character is what was triggering it.
I also found it absolutely cannot draw a young girl in fantasy armor. Almost every generation came out looking like a teenager or adult, and almost every one of them was busty with armor designed to accent her cleavage. This is no doubt because of the fantasy art on which the generator has been trained.
It did, early on, try to make child-like characters, but they inevitably came out as oppai lolis, some of which were rather disturbing. I tried to tweak the prompt to stop that from happening, but it either made no difference or hit me with the “dangerous content” warning again. The generator does not recognize operators like no or not or the minus sign. Ironically, I was trying to tone the resulting images down, but that made the problem worse.
If I didn’t ask for a specific art style, it would generate either anime-like or “photorealistic” but plasticky looking images, undoubtedly because it was drawing on a database of anime and anime-style artwork.
Some of the images are impressive in many ways, but they inevitably have the problems typical of AI art. Especially, the generator has a rough time drawing the spear going behind any part of her body. In some images, it will go behind one leg and pop out behind the other one.
Also, her armor is usually complex, but a close look shows it contains a lot of “noise” and very little rhyme or reason in its design.
I didn’t specify anything about skin tone in my prompt. For some reason, the generator appeared to have an easier time keeping track of details when it made Dynamo black.
I get the impression that the generator latches onto certain words and gives them emphasis. The first time, it apparently latched onto magical girl, or maybe just the request for fantasy armor, and generated a lot of obviously anime-inspired art, sometimes even giving an anime style without being prompted.
The second time I did this, however, it clearly latched onto the words Art Deco, and the results were unexpected and interesting. One of them is in the graphic at the top of the post, but here are some others:
Again, these have obvious AI problems. In that last one, the AI clearly got confused by the skirt, so her hips and waist are disconnected from each other.
Although I didn’t think of it myself, I like this style because it looks like the kind of propaganda posters that Urbanopolis might produce. Maybe there would be Art Deco artwork of Pretty Dynamo hanging throughout the city.
