linyq/kiwi_edit_training_data

--- task_categories: - image-to-video --- # RefVIE (Kiwi-Edit Training Data) [**Project Page**](https://showlab.github.io/Kiwi-Edit) | [**Paper**](https://arxiv.org/abs/2603.02175) | [**GitHub**](https://github.com/showlab/Kiwi-Edit) **RefVIE** is a large-scale dataset tailored for instruction-reference-following video editing tasks, introduced in the paper "**Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance**". The dataset was constructed using a scalable data generation pipeline that transforms existing video editing pairs into high-fidelity training quadruplets. It leverages image generative models to create synthesized reference scaffolds, bridging the gap in high-quality paired training data for reference-guided editing. ## Dataset Format The training metadata is provided in CSV format. The dataset supports three main training stages as described in the official repository: - **Image stage**: Includes source/target frames and instructions. - Fields: `src_video`, `tgt_video`, `prompt` - **Video stage**: Includes source/target video pairs and instructions. - Fields: `src_video`, `tgt_video`, `prompt` - **Reference-video stage**: Includes source/target video pairs, a reference image, and instructions. - Fields: `src_video`, `tgt_video`, `ref_image`, `prompt` For more details on how to use this data with the Kiwi-Edit framework, please refer to the [GitHub repository](https://github.com/showlab/Kiwi-Edit). ## Citation If you use this dataset or the Kiwi-Edit framework in your research, please cite: ```bibtex @misc{kiwiedit, title={Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance}, author={Yiqi Lin and Guoqiang Liang and Ziyun Zeng and Zechen Bai and Yanzhe Chen and Mike Zheng Shou}, year={2026}, eprint={2603.02175}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2603.02175}, } ```--- 任务类别： - 图像转视频（image-to-video） --- # RefVIE（Kiwi-Edit训练数据集） [**项目页面**](https://showlab.github.io/Kiwi-Edit) | [**论文**](https://arxiv.org/abs/2603.02175) | [**GitHub仓库**](https://github.com/showlab/Kiwi-Edit) **RefVIE** 是专为指令-参考遵循型视频编辑任务打造的大规模数据集，由论文《Kiwi-Edit：基于指令与参考引导的通用视频编辑》提出。 该数据集采用可扩展的数据生成流水线构建，将现有视频编辑样本对转换为高保真的训练四元组。它借助图像生成模型生成合成参考支架，填补了参考引导式编辑领域高质量配对训练数据的缺口。 ## 数据集格式 训练元数据以CSV格式提供。该数据集支持官方仓库中所述的三大训练阶段： - **图像阶段**：包含源/目标帧与指令，字段包括：`src_video`、`tgt_video`、`prompt`（提示词）。 - **视频阶段**：包含源/目标视频对与指令，字段包括：`src_video`、`tgt_video`、`prompt`（提示词）。 - **参考视频阶段**：包含源/目标视频对、一张参考图像与指令，字段包括：`src_video`、`tgt_video`、`ref_image`、`prompt`（提示词）。 有关如何结合Kiwi-Edit框架使用该数据集的更多细节，请参阅[GitHub仓库](https://github.com/showlab/Kiwi-Edit)。 ## 引用声明 若您在研究中使用该数据集或Kiwi-Edit框架，请引用如下文献： bibtex @misc{kiwiedit, title={Kiwi-Edit: Versatile Video Editing via Instruction and Reference Guidance}, author={Yiqi Lin and Guoqiang Liang and Ziyun Zeng and Zechen Bai and Yanzhe Chen and Mike Zheng Shou}, year={2026}, eprint={2603.02175}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2603.02175}, }