Read the original document by opening this link in a new tab.
Table of Contents
1. Introduction
2. Related Work
3. Method Preliminaries
4. Experiments
4.1. Qualitative Results
Summary
The document discusses the CONCEPTOR method for interpreting the internal representations of text-to-image diffusion models. It introduces a novel approach to decompose textual concepts into interpretable elements, revealing non-trivial structures in the representations. The method aims to provide meaningful, robust, and faithful decompositions for a wide variety of abstract, concrete, and complex textual concepts. Through experiments, it demonstrates the method's effectiveness in understanding the inner workings of text-to-image diffusion models.