For best outputs and usage:

Sampler: Euler a.

Steps: 20 ~ 40, cfg 4.5 ~ 6, CLIP skip 2.

VAE: none required (baked in).

Pos prompt: start it with "score_9, score_8_up, score_7_up, score_6_up", being "score_6_up" rather optional, remove it for a rather western cartoon look, but doing so will also also possibly sacrifice some of the model's understanding. It's recommended including "score_6_up". Also, this model can switch between 3 cartoon styles, anime, cartoon and semi-realism. To achieve this, you have "source_anime", "source_cartoon" and "source_realism" (source_realism creates semi-realism). Place the one you want in the pos and the other 2 in the neg.

Note on the pos: besides of using the source tags, you can help the model achieve the style by using keywords. For example, for source_realistic, say "photo" instead of "illustration", or say "photorealism". For anime, say "2d", or "japanese animation", ect. The model is by default set to cartoon without any particular western / eastern anime style. Also, mixing keywords (for example, source_realism + 2D) will likely create a mix of both.

Neg prompt: "score_6,score_5,score_4, busty, double cock, ugly face, child, kid, loli, underage, mutated hands, low res, blurry face, pumped body, athletic body, sticks, muscular female, bad anatomy, bad proportions, deformed, deformed anatomy, deformed fingers, cartoon, skinny, long torso, messy_hair, censored, lips, long neck" plus what you don't want to see / source tags.

Overall note: sometimes maybe the model tends to blend / mix characters / outfits; although it's not common to happen. There is an embedding included in most of my example runs, "ng_deepnegative_v1_75t", which may help with this issue.

Checkpoint Trained.