I strive for balance between the two but it's a balancing act! There's a lot of photorealistic nobodies, so this is an option - that said, I generally want them to work in those models before posting. If they don't, I make a note of it. This one, Jen, is probably the best for photorealism!

... To my never ending annoyance, this is the one model I have lost the Ref Data for, which is a shame, because I think it's the one that came out the best.

With all the showcase images and many of the better results in the model gallery though, a new dataset should be makeable, and I'll likely do one soon, upscaling where necessary.

For SDXL, while we're finally seeing embeds, they seem to be more akin to Textual Inversions than trained on reference images, but I haven't looked into it much, mainly because training at 1024x1024 is asking too much of my 3070. Once I get a dataset, I'll look into throwing them into the Civit LoRA maker. Not as ideal as a lightweight embed, but worth trying.