SCFlow 数据集
收藏资源简介:
SCFlow 数据集是一个包含 510,000 个样本的大规模数据集,旨在模拟解耦。该数据集由 51 种艺术风格和 10,000 个内容实例的全面组合覆盖组成。每个内容实例都与每种风格配对,使得模型可以学习如何将风格和内容合并,并通过反向过程进行解耦。该数据集的构建使得 SCFlow 能够通过观察风格和内容的独立变化来隐式地学习解耦,即使在没有观察到“干净”的表示的情况下也能隐式地推断出风格和内容。
The SCFlow dataset is a large-scale dataset containing 510,000 samples, designed to simulate disentanglement. It covers a complete combinatorial set of 51 artistic styles and 10,000 content instances. Each content instance is paired with every style, enabling the model to learn how to merge style and content and perform disentanglement via the reverse process. The construction of this dataset allows SCFlow to implicitly learn disentanglement by observing the independent variations of style and content, and to implicitly infer style and content even when no "clean" representations are observed.
SCFlow数据集概述
基本信息
- 数据集名称: SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models
- 发布机构: CompVis Group @ LMU Munich, Munich Center for Machine Learning (MCML)
- 作者: Pingchuan Ma, Xiaopei Yang, Yusong Li, Ming Gui, Felix Krause, Johannes Schusterbauer, Björn Ommer
- 相关论文: ArXiv论文
- 会议: ICCV 2025
数据集状态
- 当前版本: 未发布
- 计划发布内容:
- 推理代码和检查点(ckpt)
- 数据集和训练代码
更新日志
- [06.08.2025]: ArXiv论文已发布

- 1SCFlow: Implicitly Learning Style and Content Disentanglement with Flow Models慕尼黑大学计算机视觉实验室(CompVis @ LMU Munich) · 2025年



