This dataset is a video instance editing dataset released by Peking University and OpenBayes Bayesian computing in 2025. The relevant paper results are:VIRES: Video Instance Repainting via Sketch and Text Guided Generation", which aims to provide accurate annotation support for tasks such as video instance redrawing and time series segmentation.

The dataset contains:

• SA-V enhances mask annotation, adds a new masklet_continues field in the original JSON file, which can be parsed by base64 decoding and pycocotools.mask tool.

• 86k video clips, including 85k training videos and 1k evaluation videos, each video consists of 51 frames of 24 FPS, 512×512 resolution images, and is accompanied by a sequence of structure sketches and appearance text descriptions.

Video Editing Examples