>>2893273I would guess because you use multiple score tags, but without cfg they're too weak to show their style bias (or do much for quality). I mean for upscaling you don't really need prompt adherence since the original image gives it enough guidance already. So a low cfg is fine, but it also weakens any prompted styles. And pony's score tags are effectively a style the same as prompting an artist, just split into six tags.
The pony "default" CFG is around 5-6 I would say, and 8 for most other models. The less training they got on top of base SDXL, the higher you can go without things burning.