Have found that this works very well with the suggested resources. Will try to iterate on this when I get a second.

Specifically using the float LoRA will sort of trigger the AI to understand you want zero gravity and pose the characters more appropriately to the game art, otherwise they'll forever stay grounded.

Plant Milk or a similar checkpoint will do a lot to stabilize human figures. Either I didn't use high enough quality images or the general 'bagginess' of the figures makes SDXL 1.0 really confused that you don't want everything to have that same quality.

Surprisingly, the LoRA is pretty good at snapping into the isometric city style of the maps of the game, which I didn't specifically try to train it for.