𝐂𝐨𝐦𝐩𝐨𝐬𝐢𝐭𝐢𝐨𝐧𝐚𝐥 𝐯𝐬. 𝐂𝐨𝐧𝐜𝐞𝐩𝐭𝐮𝐚𝐥 𝐍𝐨𝐯𝐞𝐥𝐭𝐲 𝐢𝐧 𝐓𝐞𝐱𝐭-𝐭𝐨-𝐈𝐦𝐚𝐠𝐞 𝐌𝐨𝐝𝐞𝐥𝐬

This paper presents an empirical study of diffusion-based text-to-image models and investigates how their creative capacity can be evaluated through the novelty of generated images. Inspired by parallels to human dreaming, we investigate whether the apparent originality of these systems reflects genuine conceptual creativity or merely the recombination of patterns learned from training data. Specifically, this work aims to assess whether diffusion models produce truly novel visual representations or operate within the constraints of their learned distributions. In this work, we explore the conceptual and computational relationship between human dreaming and diffusion-based image generation models. Dreams are often understood as a recombination of stored memories into unfamiliar configurations. Although dream experiences feel novel, they are constructed from prior representations. Inspired by this analogy, our study investigates whether the generative behavior of diffusion models can be analyzed through a framework inspired by dream-like visual generation, focusing particularly on the balance between semantic alignment and visual novelty. Our work introduces a novel perspective that integrates the concept of human dreams into the evaluation of generative diffusion models. By analyzing the generated images using multimodal similarity metrics and novelty estimation techniques, we aim to investigate how diffusion models balance creativity and semantic consistency. The central research question guiding this study is: To what extent do diffusion-based generative models produce genuinely novel visual concepts, and how similar is this process to the compositional and associative mechanisms observed in human dreaming? Access the code and all related work on GitHub repo