Some images remain in memory while others fade quickly. Memorability reflects an observer-independent, intrinsic property of an image, which can be predicted by computational models. While semantic content has been proposed as the critical determinant of memorability, it is challenging to disentangle image content from visual features on display. Here, we find that the memorability of scene images transfers to short textual descriptions of those scenes, whether generated by humans or multimodal large language models. Furthermore, new images produced with generative AI from descriptions alone reproduce the memorability of original images. Together, these findings provide strong evidence that the memorability of scene images can be explained by their semantic content, regardless of visual features on display, and that language serves as an efficient and faithful compression of memory-relevant content.
Kollenda et al. (Sun,) studied this question.