Sparkle: ~140K Video Pairs for Background Replacement

Sparkle is a large-scale video background replacement dataset created by Ziyun Zeng, Yiqi Lin, Guoqiang Liang, and Mike Zheng Shou. It comprises approximately 140,000 high-quality source–edited video pairs. The dataset was last updated on 2026-05-08 and is organized into five themes.

Use Cases

• Training models for instruction-guided video background replacement based on the described source–edited pairs.
• Benchmarking video editing algorithms on a large-scale dataset of ~140K video pairs.
• Developing video generation models that can follow textual instructions to modify backgrounds.
• Researching decoupled guidance methods for video manipulation tasks.

Strengths

• Contains approximately 140,000 video pairs, indicating a substantial scale.
• Video pairs are described as high-quality.
• The dataset is organized into five distinct themes, suggesting structured diversity.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

stdKonjac (author) on Hugging Face.

Collection Method

Methodology and dataset details are referenced to an accompanying paper.

Freshness

Last updated 2026-05-08 08:24:24; freshness should be verified.