Ditto-1M
收藏资源简介:
Ditto-1M是一个大规模、高质量的开源指令-视频对数据集,包含超过一百万个高保真视频编辑示例。该数据集由Ditto框架生成,该框架融合了先进的图像编辑器和上下文视频生成器,克服了现有模型的局限性。Ditto-1M数据集旨在解决指令式视频编辑中大规模、高质量训练数据的稀缺问题,通过高效的蒸馏模型架构和智能代理的辅助,实现了数据生成的可扩展性。Ditto-1M数据集可用于训练视频编辑模型,并已在Editto模型上取得了优异的性能。
Ditto-1M is a large-scale, high-quality open-source instruction-video pair dataset containing over one million high-fidelity video editing examples. This dataset is generated by the Ditto framework, which integrates state-of-the-art image editors and contextual video generators to overcome the limitations of existing models. Designed to address the scarcity of large-scale, high-quality training data for instruction-driven video editing, Ditto-1M achieves scalable data generation through an efficient distillation-based model architecture and the assistance of AI Agents. The Ditto-1M dataset can be used to train video editing models and has achieved excellent performance on the Editto model.

- 1通过香港科技大学,蚂蚁集团,浙江大学,东北大学 · 2025年



