For best outputs:

Sampler: DPM++ 2M Karras.

Steps: 35, cfg 4 ~ 6, CLIP skip 2.

VAE: kl-f8-anime2_fp16.safetensors (optional).

Pos prompt: just what you want to see. The model listens well, but you may need big weights when it comes to solo persons, specially with a 768 x 768 resolution. To solve, change to 512 x 768. Use the LoRa "add_detail" with a weight of 1 at the end of your pos (included in my example runs).

Neg prompt: "ng_deepnegative_v1_75t, easynegative, FastNegativeV2, cross-eyed, bad anatomy, (worst quality:2), (low quality:2), nudity", being the embeddings "ng_deepnegative_v1_75t", "easynegative" and "FastNegativeV2" found in my example runs.

Checkpoint Trained.