Instruction-Guided Video Editing Dataset With 3 Million Samples

OpenVE-3M is a large-scale dataset for instruction-based video editing, containing 3 million samples. It was developed by researchers from Zhejiang University and Bytedance and released as an open-source resource in 2024.

Use Cases

• Train video editing models using instruction-text and source video pairs to generate edited outputs.
• Benchmark the performance of generative AI models on tasks specified by textual instructions.
• Develop and evaluate multi-modal models that process both video and text modalities for content manipulation.

Strengths

• Contains 3 million samples, providing a substantial scale for model training.
• Described as high-quality and open-source, facilitating academic and commercial use.
• Covers eight major editing categories, offering diverse task coverage.

Limitations

• Specific video resolution, duration statistics, and detailed column schema are not provided in the input.
• Potential geographic or content bias is unknown as spatial coverage and source material details are unspecified.

Provenance

Source

Hugging Face dataset repository 'Lewandofski/OpenVE-3M'.

Collection Method

Constructed as a dataset for instruction-guided video editing; specific gathering methodology is not detailed in the provided input.

Freshness

Dataset page was last updated on 2026-01-26, indicating recent maintenance.

Geography

Region tag includes 'us', but specific spatial coverage of the video content is unknown.