>>8651474
It also produces different loss weights, according to the training graphs, compared to the regular edm2 implementation for me. And these actually make sense.