The "NSFW" identity is there to reduce the amount of eyeballs that would otherwise be guided to this LoRA without proper caution. However, if your intent is to produce pornographic material you'll be quite disappointed. The clarification is cautionary only since the LoRA will produce nudity in various forms.

Why reality augmentation for models that already perform well?

The theory is that the training data strengthens the focus of the model by limiting its generic range, which aught to make the model produce images that are more coherent for a particular purpose. In this case the focus is on people so that there are fewer spare parts, reducing body horror, i.e. extra limbs etc.

This is why adding almost any LoRA can have a desirable effect on a base model in a way that was not directly intended.

This was trained on 407 generic people pictures, mostly women and that's not exactly a biased thing because the base model was probably trained on, as far as people, mostly women so by balancing with the same occurrences we hope that we guide the pathways away from the area that brings the most confusion, which may seem a bit confusing but it is a sound concept.

This has a wide weight range tested between 0.25 and 1.2.