RoughCut-Bench

A Benchmark for Automatic Video Rough-Cut with Large Vision-Language Models

Accepted to KDD 2027

Ruisi Zhao1,2You Li1,2 Sihan Chen2,*Aosong Cheng2,3 Chixiang Ma2
Sen Pei2Zhen Yang2 Hehe Fan4Dongliang He2 Yi Yang1,†
1ReLER, CCAI, Zhejiang University 2ByteDance 3Peking University 4School of Artificial Intelligence, Zhejiang University
*Project Leader †Corresponding Author

Overview

Evaluating executable rough-cut plans

Video rough-cut organizes raw media into an initial cut through highlight detection and selection, source-video ordering, and spoken-content refinement. RoughCut-Bench studies this compositional creation task through visually driven video cases and speech-driven ASR cases.

The benchmark combines annotation-based metrics with human-aligned assessment of rendered cuts. Proposed agentic workflows guide representative LVLMs to produce structured cut plans from the input materials.

101Visual cases
831Raw videos
1,476Highlight annotations
39Cases with requirements
84Speech cases
Overview of the RoughCut-Bench settings, proposed agentic workflows, and evaluation protocols.
Benchmark settings, agentic workflows, and evaluation protocols.
Dataset distributions for visual scenarios, video duration, and speech cleanup annotations.
Dataset composition and annotation distributions.

Leaderboard

Performance across rough-cut dimensions

Click a metric header to rank models. No cross-setting aggregate is used.

Gemini GPT Qwen Other

Case Explorer

Inspect the inputs, references, and outputs

All public examples are privacy reviewed; visual media are muted.

Citation

Citation

Citation information will be added after publication.