Interesting, The model says 93% accurate but in my testing it was less than 50% accurate. From what I can tell it mostly seems to base its assessment on ow much of the picture is white skin tone. It does a very poor job of analyzing shapes and structures in the picture.

Some examples of its failure that I found, was close up pictures of faces often get flagged as porn. Probably because these pictures were not in public learning models. On the other hand pictures that are mostly clothed and have neutral backgrounds but there is a bare buttocks exposed in a lewd manner get flagged as neutral. They barely rank on the porn or sexy percentage.