GOKU-2M: A Million-Scale Instruction-Based Video Editing Benchmark

GOKU-2M is a large-scale, unified dataset for instruction-based video editing, covering 10 editing tasks. Each sample provides a source video, an edited target video, and one or more natural-language instructions describing the edit. The dataset was created by 'bigfacing' and was last updated on Hugging Face in July 2026.

Use Cases

• Training video editing models based on natural language instructions.
• Benchmarking model performance across 10 different video editing tasks.
• Developing multimodal AI that understands and executes edits from text descriptions.
• Studying the relationship between textual instructions and visual transformations in video.

Strengths

• The dataset is described as 'large-scale' and 'million-scale', indicating substantial size.
• It covers 10 distinct video editing tasks, suggesting breadth in editing types.
• Each sample includes a source video, target video, and natural language instruction, providing a complete training triplet.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• The dataset is split across multiple repositories due to storage constraints, which may complicate access.

Provenance

Source

huggingface

Freshness

Last updated 2026-07-01 13:53:39; freshness should be verified.