FastVideo

Four FastH3 4-Step Preview v1 T2VA checkpoints from FastVideo: VSA and dense attention, with data-free and synthetic-data recipes.

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