FastVideo

models trained with video sparse attention: https://arxiv.org/abs/2505.13389 and distillation