For best outputs:

Sampler: DPM++ 3M SDE / DPM++ 2M Karras for extra photorealism.

Steps: around 30, cfg 4 ~ 7, CLIP skip 2.

VAE: sdxl_vae.safetensors.

Model doesn't need big or long prompts, neg can be left empty or simply "underdetailed, blur, blurry, worst quality".

This model is good at sticking to the prompt while also looking a bit different than most other models.