metadata

license: cc-by-nc-4.0
task_categories:
  - video-to-video

LongV-EVAL: A Benchmark for Long Video Editing Evaluation

Paper

LongV-EVAL is a benchmark dataset designed for evaluating text-driven long video editing methods. It consists of 75 high-quality videos, each approximately one minute long, covering diverse domains such as landscapes, people, and animals. The dataset provides meticulously annotated editing prompts for three aspects: foreground, background, and style, enabling comprehensive evaluation of editing quality, temporal consistency, and semantic alignment.

Dataset Structure

The dataset is organized into four folders:

videos/: Contains 75 MP4 files of source videos (original unedited videos).
foreground/: Includes 75 text files with prompts focusing on foreground object editing (e.g., changing object attributes or replacing objects).
background/: Includes 75 text files with prompts for background modification (e.g., altering scene context or tone).
style/: Includes 75 text files with prompts for artistic style transfer (e.g., applying styles like Van Gogh, watercolor, or Picasso).