This paper concerns the structure of learned representations in text-guided generative models, focusing on score-based models. Here, we focus on the idea that concepts are encoded as subspaces (or directions) of some representation space. We develop a mathematical formalization of this idea.Using this formalism, we show there's a natural choice of representation with this property, and we develop a simple method for identifying the part of the representation corresponding to a given concept. In particular, this allows us to manipulate the concepts expressed by the model through algebraic manipulation of the representation. We demonstrate the idea with examples text-guided image generation, using Stable Diffusion.

翻译：本文关注文本引导生成模型中学习到的表示结构，重点关注基于得分的模型。我们聚焦于概念被编码为某种表示空间的子空间（或方向）这一思想，并为其建立数学形式化框架。利用这一形式化方法，我们证明存在具有该属性的自然表示选择，并开发出一种识别表示中对应给定概念部分的简单方法。特别地，这使得我们能够通过对表示进行代数操作来操控模型所表达的概念。我们通过使用Stable Diffusion的文本引导图像生成实例来演示这一思想。