[D] Correct me if I'm wrong, use KL divergence for NLP, and MMD for CV. Both are measuring the similarity/distance of two distribution

I'm doing instance selection now, and after a quick research, I found fewer works are using MMD rather than KL divergence. Is there a preference for choosing a distance metric in these two fields? What is the reason behind this? THX

distribution divergence fields found instance machinelearning measuring nlp research