I am extremely skeptical, because this looks so good. To my eyes the images produced look orders of magnitude better than what current software is capable of producing. Let alone from multiple perspectives with absolutely perfect coherence. What tool was used?
Stuff like Dall-e is still far from producing anything like that. Those are absolutely perfect idealized/stylized human proportions with minimal weirdness. The only thing that even hints at neural network generation (at a glance at least) are the eyes in the second model sheet, but even that could be reasonably argued to be style.
I could fully believe this is from a specialized art tool akin to some sort of a 2D version of Metahumans [1] but I'd be floored if this was actually generated from a generic neural network style tool.
I had the same reaction when I tried midjourney but this looks very believable.
https://www.midjourney.com/app/feed/all/
I don't think generic AIs are the right approach anyway, I think we're better off with precise procedural models with a lot of knobs and some knowledge of rules (eg. metahumans).
Maybe you can put an AI in front of it to generate what you want using the procedural models, but pixel output is just not good enough for games.
I've been using https://github.com/xinntao/Real-ESRGAN to increase AI renders resolution, it works very very well on some styles, you can easily x4 or x8 the resolution if you know what you're doing and have some photoshop knowledge for the cleanup