作者：Nupur Kumari Bingliang Zhang Sheng-Yu Wang Eli Shechtman Richard Zhang Jun-Yan Zhu
大规模的文本到图像扩散模型可以生成具有强大合成能力的高保真图像。然而，这些模型通常是在大量互联网数据上训练的，这些数据通常包含版权材料、授权图像和个人照片。此外，他们被发现可以复制各种在世艺术家的风格或记忆精确的训练样本。我们如何在不从头开始制作模型的情况下删除这些受版权保护的概念或图像？为了实现这一目标，我们提出了一种有效的方法来消除预训练模型中的概念，即防止目标概念的产生。我们的算法学习将目标风格、实例或文本提示的图像分布与锚概念对应的分布相匹配。这将阻止模型在给定其文本条件的情况下生成目标概念。大量实验表明，我们的方法可以成功地防止在预处理时产生烧蚀概念
Large-scale text-to-image diffusion models can generate high-fidelity imageswith powerful compositional ability. However, these models are typicallytrained on an enormous amount of Internet data, often containing copyrightedmaterial, licensed images, and personal photos. Furthermore, they have beenfound to replicate the style of various living artists or memorize exacttraining samples. How can we remove such copyrighted concepts or images withoutretraining the model from scratch? To achieve this goal, we propose anefficient method of ablating concepts in the pretrained model, i.e., preventingthe generation of a target concept. Our algorithm learns to match the imagedistribution for a target style, instance, or text prompt we wish to ablate tothe distribution corresponding to an anchor concept. This prevents the modelfrom generating target concepts given its text condition. Extensive experimentsshow that our method can successfully prevent the generation of the ablatedconcept while preserving closely related concepts in the model.
论文链接：http://arxiv.org/pdf/2303.13516v1
更多计算机论文：http://cspaper.cn/