For best outputs and usage:

Sampler: Euler a.

Steps: 30, cfg 7.5, CLIP skip 2.

VAE: sdxl_vae.safetensors (mandatory).

Pos prompt: start it with "score_9, score_8_up, score_7_up, source_anime". The model listens quite well and is able to stand certain level of weighting to guide it towards what you want to see.

Neg prompt: "score_4, score_5, score_6, zombie, bad fingers, bad hand, bad feet, extra digits, (interlocked fingers, badly drawn hands and fingers, anatomically incorrect hands), signature, (lipgloss, lipstick:0.7), lowres, missing, worst quality, jpeg artifacts, low quality, unfinished, displeasing, oldest, signature, jpeg artifacts, blurry, signature, watermark".

Checkpoint Merge.