KoSum

Abstract

This project proposes a framework for practical video editing guidance. Existing highlight detection and video summarization models remain limited for real-world editing, where creators need actionable guidance rather than simple cut-based automation. We construct KoSum, a Korean YouTube benchmark for video summarization, consisting of recent videos from 2024 to 2026 collected through a structured protocol across 14 fine-grained content categories (e.g., entertainment and cooking) and annotated with triple-modality signals (e.g., visual, audio, and text) and Most-Replayed signals. Using a triple-modality model, we analyze highlight regions through modality attention patterns and feature-level statistics, capturing editing-related multimodal cues (e.g., shot transitions, motion dynamics, subtitle density, and audio novelty). Based on these analyses, we propose a data-analysis-conditioned prompting framework that augments user prompts with category-specific and statistically grounded editing cues. The proposed framework outperforms the baseline by 95 points in total score and 3.17 points in average score, demonstrating strong effectiveness in generating practical and structured editing guidelines. Our framework connects viewer engagement signals, multimodal analysis, and practical editing support for real-world creators. The project page is available at https://iontail.github.io/kosum/.