Trained on the same dataset initially used for the Flux LoRA, results with SDXL are equally charming and sometimes even a bit more authentically handmade-feeling. Flux brings better prompt comprehension, but the SDXL model costs a lot less in compute.

Creating scenes of miniatures - you don't always need any key words, but it reinforces the style to start with 'diorama miniatures view of ...' or 'overhead diorama view of ...'

You can get close to this look with extensive prompting, but the LoRA helps to keep texture and aesthetic more consistent while simplifying your prompts.