MUFASA: A Multi-Layer Framework for Slot Attention

A new framework called MUFASA improves how machine-learning models segment objects in images without human supervision by tapping into information previously ignored inside vision transformers, according to research posted on arXiv [1]. Unsupervised object-centric learning, or OCL, aims to break down visual scenes into distinct entities. A widely used technique, slot attention, represents individual objects as latent vectors known as slots. Existing methods derive these slot representations exclusively from the final layer of a pre-trained vision transformer, or ViT, discarding semantically rich information encoded in earlier layers [1][2]. Sebastian Bock and colleagues propose MUFASA, a lightweight plug-and-play framework designed to address this gap. The model computes slot attention across multiple feature layers of the ViT encoder and then applies a fusion strategy to aggregate the resulting slots into a unified object-centric representation [1][2]. When integrated into current OCL methods, MUFASA improved segmentation results across multiple datasets and set a new state of the art, the authors report. The framework also accelerated training convergence while adding only minor inference overhead [1][2]. The paper was submitted to arXiv on February 7, 2026, with a revised version posted on June 17, 2026. The submission file size for the initial version was 21,463 KB, shrinking slightly to 21,460 KB in the update [1]. arXiv, where the research appeared, is an open-access repository of electronic preprints that are moderated but not peer-reviewed. It hosts papers across disciplines including computer science, physics, and mathematics. As of November 2024, the repository was receiving about 24,000 new articles per month [6]. The paper’s abstract page features arXivLabs, a framework launched in 2020 that enables community collaborators to develop and share experimental tools directly on the site, such as citation explorers and code finders [4][5].