Introduces a way for diffusion models to generate a single, sharp 'mental average' of a concept rather than blurry pixel-wise averages.
April 1, 2026
Original Paper
Diffusion Mental Averages
arXiv · 2603.29239
The Takeaway
DMA operates in the semantic latent space of the model rather than the image space, allowing for the extraction of 'prototypical' visual summaries. This provides a new lens for researchers to audit model biases and understand how abstract concepts are internally represented.
From the abstract
Can a diffusion model produce its own "mental average" of a concept-one that is as sharp and realistic as a typical sample? We introduce Diffusion Mental Averages (DMA), a model-centric answer to this question. While prior methods aim to average image collections, they produce blurry results when applied to diffusion samples from the same prompt. These data-centric techniques operate outside the model, ignoring the generative process. In contrast, DMA averages within the diffusion model's semant