CI-VID is a large-scale dataset designed to advance coherent multi-clip video generation. Unlike traditional text-to-video (T2V) datasets with isolated clip-caption pairs, CI-VID supports text-and-video-to-video (TV2V) generation by providing over 340,000 interleaved sequences of video clips and rich captions. It enables models to learn both intra-clip content and inter-clip transitions, fostering story-driven generation with… See the full description on the dataset page:
https://huggingface.co/datasets/BAAI/CI-VID.
