Large text-to-image models can generate "novel" images, but it is difficult to determine which training images are responsible for the appearance of a generated image. This paper proposes a method to evaluate data attribution in these models by customizing them towards exemplar objects or styles. The authors create a dataset of exemplar-influenced images to evaluate different attribution algorithms and feature spaces. They also show that training on this dataset can generalize to larger exemplar sets and assign soft attribution scores.