/vt/ - Virtual Youtubers

I humbly beseech your expertise /vtai/, I'm having a really bad issue with deep fried faces!

Recently did a clean install of a1111 after not proompting for a year. I'm using Autismmix with a VAE that's just called sdxl.vae.safetensors (I kind of hope this is the culprit, idk what people use for pony models), and I've only been using LoRAs made for/with for Pony.
I'm prompting with DPM++2 SDE, Karras and 28 steps for sampling.
I'm using CFG scale 7, and I'm hires-ing with 4xAnimesharp at 0.7 denoising strength. I'm also using clip skip 2.
I have only been using the aspect ratios Pony was supposedly trained by using.

Through my Botan proompting I've tried using the holo-ponyxl-te-only-128-lion-v2 LoRA and the Pony lora for Botan in the OP.
I've also used two different "thiccwithaq"/ Nyantcha style LoRAs and a Mogudan one.
Without any LoRAs and before upscaling my faces are already deepfried. After upscaling they remain deepfried. It's especially the eyes that are bad.
Literally everything else besides the eyes and the faces is PERFECT. No weird artifacts, no nothing.

Really not sure what to do, hope it's somehow just a matter of switching VAE.
Have some of my best shots of BEEG butt Botan being analyzed as a thanks.

https://files.catbox.moe/o5vq4t.png
https://files.catbox.moe/duls7v.png
https://files.catbox.moe/70ltt9.png
https://files.catbox.moe/0l9bqg.png
https://files.catbox.moe/bedu1q.png
https://files.catbox.moe/p9zq8e.png

It's very late, but I'll stay up a while to look for any tips, and I'll look at any replies tomorrow too.