>>107046709
Yeah, it's the VAE I use. While it enhances my LoRA that was trained with an EQ VAE, it "sharpens" the output just enough to pull that grid out of SRPO/other LoRAs. The guy suggests not using his EQ VAE for inference, but the color is much more vibrant and my LoRA holds up much better (even when the face is very small!) that I prefer it to the default even with the extra noise.