Sign in to view source links and access this dataset
Description
Sparkle is a large-scale dataset containing approximately 140,000 high-quality source-edited video pairs for video background replacement. It was created by Ziyun Zeng, Yiqi Lin, Guoqiang Liang, and Mike Zheng Shou and is hosted on Hugging Face. The dataset is organized into five distinct themes, as detailed in the associated research paper.
Use Cases
Train models for instruction-guided video background replacement based on the described source-edited pairs.
Benchmark video editing algorithms using the dataset's five thematic categories.
Develop video synthesis tools leveraging the high-quality paired examples.
Research decoupled guidance methods for video manipulation as referenced in the paper title.
Strengths
Contains approximately 140,000 video pairs, indicating a substantial scale.
Described as high-quality source-edited pairs, suggesting curated content.
Organized into five themes, providing structured variation.
Limitations
Column-level documentation is absent; field semantics must be inferred after download.
Row count is unknown, which may limit suitability assessment.
Freshness should be verified as the last update timestamp is in the future (2026-05-08).
Provenance
Source
Hugging Face user stdKonjac, associated with academic authors Ziyun Zeng, Yiqi Lin, Guoqiang Liang, and Mike Zheng Shou.
Collection Method
Likely created for research on instruction-guided video background replacement, as per the paper title.
Freshness
Last updated 2026-05-08 08:20:41
License is unknown; terms of use must be verified before application.