In-house Magi-1 autoregressive video model, GAGA-1 audio-video synchronized generation model, and MAGI-2 Previ