CI-VID

📄 CI-VID: A Coherent Interleaved Text-Video Dataset 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 strong temporal and visual coherence. We also introduce a comprehensive evaluation suite including **human**, **VLM-based**, and **similarity-based** assessments. In addition, we split CI-VID into a training set (334k samples) and a test set (8k samples), enabling users to develop and evaluate their own metrics using the test set. 🔗 📃 Paper 🔗 📦 Download Train Samples (334k) 🔗 📦 Download Test Samples (8k) 📦 Download Videos (6.9T): CI-VID/CI-VID_download_all_videos.sh 🗂️ Samples of CI-VID 📁 **Provided Files** `CI-VID_samples_for_visualization/` This part of the repository contains samples extracted from CI-VID to better illustrate the dataset’s structure and characteristics. 📊 Quantitative Evaluation for CI-VID This part of the repository contains the quantitative evaluation resources for the CI-VID dataset, including evaluation code, prompts, visualizations, and sample annotations. We provide three complementary evaluation approaches: 🔍 Overview We propose three evaluation protocols: 1. **Human Evaluation** 2. **VLM-based Evaluation** 3. **Similarity-based Evaluation** 👥 1. Human Evaluation 📁 **Provided Files** `human_evaluation/prompts.jsonl` → *Prompts used for evaluation.* `human_evaluation/visual_contrast/` → *Visualizations for human evaluation (1,000 prompts). 💡download Human evaluation is based on 1,000 automatically generated prompts (Seeded with keywords from VBench), with each describing 6 scenes composing a coherent multi-scene narratives. **Models Compared**: