I saw someone said that the recognition of "yoga pants" was not good enough, and you replied that you would add new training. So I shamelessly hope that you can add more training for "2D" (anime characters) and the like. At present, when recognizing the breasts of some relatively abstract animated human structures, it is often found that "censorship" does not work. This can also be found when looking at many pictures produced by AI. I wonder if you have the time and energy to do this? (The above is the Chinese translation of Google, and the expression may not be very clear)

I don't think it's really the yoga pants that were the problem. It's the poses that they are making. I love yoga and especially gymnastics pics like this and as you can see the pose is def unusual lol.