We　propose　an　efficient　real-time　character　garment　animation　simulation　method　based　on　deep　learning.　Given　a　character　model　and　garment　model,　we　create　a　database　with　character　animations　and　corresponding　garment　animations　for　training.　For　garment　mesh　with　many　vertices,　we　use　an　autoencoder　to　extract　low-dimensional　features　in　subspace,　which　greatly　reduces　computational　cost.　Then　we　build　an　animation　inference　network　designed　based　on　VRNN.　The　state　of　the　previous　frame　and　the　motion　of　the　character　are　used　together　to　update　hidden　state.　At　runtime,　input　the　character　animation　to　the　animation　inference　model　to　get　the　garment　feature,　and　decode　it　into　the　vertex　position　of　the　garment　model.　This　method　aims　at　the　specific　issue　of　character　garment　animation　and　observes　its　high　correlation　with　character　motion.　It　can　calculate　the　vertex　animation　of　a　complex　garment　model　in　a　few　milliseconds.　©　2021　IEEE.