EvolvED: Evolutionary Embeddings to Understand the Generation Process of Diffusion Models

Abstract Diffusion models, widely used in image generation, rely on iterative refinement to produce images from noise. Understanding this data evolution supports model development and interpretability, yet is challenging due to its high‐dimensional, iterative nature. Prior works often focus on static or instance analyses, missing the iterative and holistic aspects of the generative space. While dimensionality reduction can visualise image evolution for some instances, it does not preserve the iterative structure. To address these gaps, we introduce EvolvED, a method that presents a holistic view of the iterative generative process in diffusion models. EvolvED goes beyond instance analysis, leveraging predefined analysis goals to streamline generative space exploration. User‐defined prompts aligned with these goals extract intermediate images, preserving the iterative context. Relevant feature extractors are used to trace the evolution of key image attributes, addressing the complexity of high‐dimensional outputs. Central to EvolvED is a novel evolutionary embedding algorithm, explicitly encoding iterations while preserving semantic and evolutionary relations of these encoded representations. It clusters semantically similar elements via a t‐SNE loss per iteration, introduces a displacement loss to represent iterations in distinct predefined spatial regions, and an alignment loss for continuity across iterations. This embedding is presented as rectilinear and radial layouts. We apply EvolvED to models like GLIDE and Stable Diffusion, demonstrating its ability to provide valuable insights into the generative process.