Reproduce “VC-Bench: Pioneering the Video Connecting Benchmark with a Dataset and Evaluation Metrics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhiyu Yin, Zhipeng Liu, Kehai Chen, Lemao Liu, Jin Liu, Hong-Dong Li, Yang Xiang, Min Zhang
The paper introduces "Video Connecting (VC)" as a novel task to generate smooth, continuous video content between two distinct video clips, addressing a critical gap in continuity-aware video generation. The authors propose VC-Bench, a comprehensive benchmark comprising a diverse dataset of 1,579 videos and a three-dimensional evaluation framework with 9 metrics, demonstrating its robustness and alignment with human judgment.
Key insights
- Connecting distinct video segments presents unique challenges related to information complexity, content difference, and temporal consistency that current models struggle with.
- Adapting diffusion transformer architectures for VC by mapping start/end clips to latent space and using SLERP for blending features provides a baseline but highlights limitations in maintaining consistency and smoothness.
- Providing richer spatiotemporal conditioning (longer video clips rather than just first/last frames) significantly improves the quality of generated connections.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.