Generate video modalities or predict future frames from any input