Your ControlNet seems a tad too loose and you might have a too high CFG.
Try without LORAs (or only ones you deem absolutely necessary and keep their influence low).
Use Canny for ControlNet and maybe reduce the denoising a bit.
Doing that (also no LORAs, no textual inversion etc, just "pure prompting") and using the latest CyberRealistic Pony I get pretty consistent results with these prompts:
Positive:
extremely realistic photo of young teenage girl, very busty, extremely giant tits, necklace, pink choker with bow, highly detailed, dark blue eyes, thin eyebrows, one hand covering mouth, other hand lifting back of dress, pink dress with flower pattern, short reddish brunette pigtails, simple hair clip, smooth shiny skin, necklace, bracelet, BREAK, filthy room with wood panels, wooden table on small carpet, sofa, air vent next to ceiling, glowing flat screen
Negative:
painted cheeks, wrinkled skin, protruding chin, anime, manga, doll eyes, dead eyes, bad anatomy, missing fingers
The room's still too clean, but then again backgrounds are most Pony models' weak point, sadly:
View attachment 4270736
View attachment 4270594
Mixing in iCatcher Realistic or similar (via Refiner) helps. But then you have to put more work into ADetailer.
And at least to me it wasn't worth it.
Trying to recreate the hand pose even more exactly (biting the finger) is super tough and I couldn't really get it right. Also adding "freckles" to the prompt and using CyberRealistic XL for ADetailer's face:
View attachment 4270714
I hope some of that helps you getting the desired results.