MemBench
收藏资源简介:
MemBench是由韩国科学技术院的研究团队开发的第一个用于评估扩散模型图像记忆缓解方法的基准数据集。该数据集包含4500条记忆图像触发提示,分别适用于Stable Diffusion 1和2模型。数据集的创建旨在通过严格的指标评估缓解方法在触发提示和一般提示上的表现,以确保在实际应用中有效解决记忆问题。MemBench的应用领域主要集中在图像生成和隐私保护,旨在解决扩散模型在特定提示下重复生成训练数据中图像的问题。
MemBench is the first benchmark dataset developed by a research team from the Korea Advanced Institute of Science and Technology (KAIST) for evaluating image memorization mitigation methods for diffusion models. It contains 4,500 memorized image trigger prompts that are respectively compatible with Stable Diffusion 1 and 2 models. The dataset was designed to evaluate the performance of mitigation methods on both trigger prompts and general prompts through rigorous metrics, so as to ensure that the memorization problem can be effectively addressed in real-world applications. The primary application domains of MemBench are centered on image generation and privacy protection, with the core objective of resolving the issue where diffusion models repeatedly reproduce images from their training data when presented with specific prompts.

- 1MemBench: Memorized Image Trigger Prompt Dataset for Diffusion Models韩国科学技术院 · 2024年



