Huh, those look pretty good. I could easily see the Vampire Twilight ones being the cover for some edgelord FiM Fiction site story, lol.
Hmmm... two experiments to test the bar?
"Star Trek Borg Twilight Sparkle" (or Princess Cadence) and "Doctor Who Cyberman".
AI collectively seem to have trouble with the Borg and Cybermen in prompts for some reason, so it's a good measure test, I think.
Niji just spat out a generic Twilight Sparkle Cyborg:
MJ V5.1 puts out similar disappointing results to what it used to:
V5.1 still does off-brand Cybermen:
The accuracy of a generated image depends entirely on how many examples there are in the training data. The pics of Imperial Stormtroopers in the training data vastly outnumber Cybermen, which is why I can very easily have it do a stormtrooper eating pizza with his helmet on, and it comes out pretty much spot-on:
Another thing is that you really have to be specific with the prompts, because the models are trained on images from many different eras in science fiction, so, for instance, if you put in "Star Trek Borg" without anything else, you get something hyper-modern and disgustingly detailed:
Now, that looks really cool, but it doesn't look
right. So, let's try that again. This time, the prompt is "
star trek borg, 90s style costume with cheesy hoses and cabling done up to look like implants":
Still not perfect, but
way closer to the actual thing.
Both Borg and Cybermen are pretty detailed character designs. They have lots of edges and visual noise. That's hard for AI image generators to replicate effectively. It's almost like camouflage.