For best outputs: Sampler: DPM++ 2M SDE Exponential Steps: 20, cfg 7, CLIP skip 1. VAE: sdxl_vae.safetensors. Pos prompt: avoit words like "masterpiece, cinematic, volumetric" ect or the model will start creating semi-realism instead of anime. Keep the pos short, no big weights. Neg prompt: this one can do "bad quality, worst quality, easynegative, child, kid, loli, underage, realistic, volumetric" being the "easynegative" embedding optional although recommended (included in my example runs).