I went and counted hands from
>>49236301Each set had 99 images total.
>No HLL-S-222 images with good hands
>HLL-S-216 images with good hands
>HLL-S-2 with badhandv4 negative embedding27 images with good hands.
For the purpose of this test, the criteria was "The hands have the correct shape and the correct number of fingers"
It seems that hll-s-2 is somewhat more likely to give weird hands, but I don't know if 300 images is a good sample size. Either way, the badhandv4 embedding does more to fix hands than hll-s-2 does to mess them up.
