Goku: A Million-Scale Universal Dataset and Benchmark for Instruction-Based Video Editing
Sen Liang ⋆, Cong Wang ⋆, Zhentao Yu, Fengbin Guan, Zhengguang Zhou, Teng Hu, Youliang Zhang, Yuan Zhou, Xin Li, Qinglin Lu, Zhibo Chen † · 2026
Read Goku: A Million-Scale Universal Dataset and Benchmark for Instruction-Based Video Editing by Sen Liang ⋆, Cong Wang ⋆, Zhentao Yu, Fengbin Guan, Zhengguang Zhou, Teng Hu, Youliang Zhang, Yuan Zhou, Xin Li, Qinglin Lu, Zhibo Chen † (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2606-30599.txt.
Description
Sen Liang and ten colleagues built Goku, a dataset for editing videos by written instruction. Existing datasets mostly cover simple appearance changes, such as swapping an object's colour. Real creative work also needs structural edits, like steering how a subject moves. The 2026 dataset answers with two million high-quality editing pairs whose instructions match the changes shown.
A synthesis pipeline breaks complex edits into controllable sub-problems, and progressive filtering checks data quality throughout. On top of the data sits Goku-Edit, a model using a multimodal language model to read instructions. Its two-branch design gives structural control to a mask branch and leaves appearance rendering to the main branch. A companion benchmark, Goku-Bench, holds one thousand human-verified cases and seven editing-specific metrics. There, Goku-Edit improved instruction following by up to eight percent over open-source rivals.
Catalog record
| CREATOR | Sen Liang ⋆, Cong Wang ⋆, Zhentao Yu, Fengbin Guan, Zhengguang Zhou, Teng Hu, Youliang Zhang, Yuan Zhou, Xin Li, Qinglin Lu, Zhibo Chen † |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 41 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2606.30599