UMO数据集
收藏资源简介:
UMO数据集是字节跳动智能创作实验室UXO团队开发的,旨在支持多身份保持的可定制数据集,包含合成和真实部分的多参考图像。数据集通过从长视频中检索每个身份的帧来构建,并使用严格的相似性过滤来确保身份的准确性。该数据集支持UMO框架的有效训练,UMO框架通过多对多匹配范式,将多身份生成重新定义为全局分配优化问题,并通过强化学习在扩散模型上进行,以最大程度地提高身份一致性并减少身份混淆。
The UMO Dataset was developed by the UXO Team from the Intelligent Creation Lab of ByteDance. It is a customizable dataset intended to support multi-identity preservation applications, and contains multi-reference images spanning both synthetic and real-world domains. The dataset is constructed by retrieving frames corresponding to each identity from long-form videos, with strict similarity filtering applied to ensure the accuracy of individual identities. This dataset facilitates effective training of the UMO framework, which redefines multi-identity generation as a global assignment optimization problem via a multi-to-many matching paradigm, and optimizes diffusion models through reinforcement learning to maximize identity consistency and reduce identity confusion.




