1 paper
Xi Lin, Feihong Zhang, Yulong Shi +6
Jointly generating future video and actions has become a standard recipe for world-action models, and the strongest systems denoise the two streams on separate schedules: actions a…