VC-Bench: Pioneering the Video Connecting Benchmark with a Dataset and Evaluation Metrics

TL;DR

This work defines Video Connecting as generating a coherent transition between two video clips, and introduces VC-Bench, a dedicated dataset and a nine-metric evaluation framework across Video Quality, Start-End Consistency, and Transition Smoothness. The authors adapt a Diffusion Transformer-based approach with latent-space mapping and SLERP conditioning to tackle VC, and evaluate six open-source models, revealing strong results for some models but clear gaps in cross-scene continuity and fluency. The VC-Bench dataset provides 1,579 high-quality videos across 15 categories, with automated and human validation confirming alignment with human preferences, and the work highlights practical applications in filmmaking, social media, and VR/AR. Overall, VC-Bench establishes a standardized, multi-faceted benchmark to drive progress in continuity-aware video generation and informs future research on longer sequences and closed-model evaluation.

Abstract

While current video generation focuses on text or image conditions, practical applications like video editing and vlogging often need to seamlessly connect separate clips. In our work, we introduce Video Connecting, an innovative task that aims to generate smooth intermediate video content between given start and end clips. However, the absence of standardized evaluation benchmarks has hindered the development of this task. To bridge this gap, we proposed VC-Bench, a novel benchmark specifically designed for video connecting. It includes 1,579 high-quality videos collected from public platforms, covering 15 main categories and 72 subcategories to ensure diversity and structure. VC-Bench focuses on three core aspects: Video Quality Score VQS, Start-End Consistency Score SECS, and Transition Smoothness Score TSS. Together, they form a comprehensive framework that moves beyond conventional quality-only metrics. We evaluated multiple state-of-the-art video generation models on VC-Bench. Experimental results reveal significant limitations in maintaining start-end consistency and transition smoothness, leading to lower overall coherence and fluidity. We expect that VC-Bench will serve as a pioneering benchmark to inspire and guide future research in video connecting. The evaluation metrics and dataset are publicly available at: https://anonymous.4open.science/r/VC-Bench-1B67/.

Figures (9)

• Figure 1: VC-Bench Overview. We introduce VC-Bench, a tailored benchmark for the novel Video Connecting task. We provide a precise definition of this task and adapt several open-source video generation models to support it. To establish a systematic and comprehensive evaluation framework, we curate a high-quality dataset spanning 15 categories. We develop 9 automated metrics to evaluate model performance across three critical dimensions: Video Quality, Start-End Consistency, and Transition Smoothness. Human subjective validation further confirms the framework’s alignment with human preferences.
• Figure 2: Video Score on VC-Bench. Based on VC-Bench, we evaluate the performance of 6 open source models on the Video Connection task. The experimental results can be seen in Table \ref{['Table2:Video Score on VC-Bench.']}
• Figure 3: Dataset Construction. We outline the construction proces s of the our dataset: Web Data Crawling and Classification: Organized into 15 major categories and 72 subcategories. Data Filtering: Selected high-quality test data through aesthetic scoring, dynamic scoring, and periodic detection. Scene Detection: Employed PySceneDetect for scene segmentation. Caption Generation: Utilized Qwen2-VL to generate high-quality caption.
• Figure 4: Distribution of Curated Data. We analyzed the dataset, including category statistics, video duration statistics (24fps), caption length statistics, and aesthetic score statistics. Statistical analysis demonstrates that our dataset exhibits high quality, diversity, and complexity, meeting the rigorous requirements for benchmark testing.
• Figure 5: Instruction text and screenshots for the subjective human evaluation.