RefVideo-6M: A Reliable Reference-Based Dataset for Instructional Video Editing

About The Dataset

Built for reliable, reference-aware instructional video editing.

RefVideo-6M is a curated benchmark that covers global editing, local editing, and controllable editing, alongside a compact set of reference-based image editing tasks designed to strengthen reference-guided video generation and editing.

Dataset Statistics

Statistics of RefVideo-6M.

The dataset includes 26 editing tasks and 6 image editing tasks. The complete statistics are shown below.

Dataset Construction

Construction pipeline of RefVideo-6M.

We train four experts in addition to eight open-source experts. In our dataset, most tasks reverse the roles of the original and edited videos to construct reliable video pairs.

Model Overview

Overview of RefMoT.

RefMoT adopts a Mixture-of-Tokens design to efficiently incorporate reference information, reducing training cost while improving reference-guided editing quality and preserving strong generalization.

Visualization

Task Gallery

Browse the dataset by media type or task category, and click any sample to inspect it in a larger preview.

Citation

Cite RefVideo-6M

Use the button to copy the BibTeX entry for this work. The full citation is provided below for convenience.