We propose an adaptive layer reuse technique that dynamically reuse intermediate feature across adjacent denoising steps to enable efficient inference of text-to-video generation models

VideoUFO is the first dataset curated in alignment with real-world users’ focused topics for text-to-video generation.