🏢 MMLab, the Chinese University of Hong Kong

DiTCtrl achieves state-of-the-art multi-prompt video generation without retraining by cleverly controlling attention in a diffusion transformer, enabling smooth transitions between video segments.