Unlocking Diffusion Hierarchies: Adaptive Timestep Selection for Zero-Shot Segmentation

A new method for zero-shot image segmentation adaptively selects denoising timesteps for each pixel, moving beyond the single-timestep approaches that dominate current diffusion-based techniques, according to a paper submitted to arXiv on 14 Jun 2026 [1]. The work targets a known limitation in zero-shot segmentation using text-to-image diffusion models such as Stable Diffusion. Existing methods extract features at one static timestep, forcing a trade-off between spatial resolution and contextual information [1][2]. The authors propose two changes. First, they introduce Contextual Similarity Maps that fuse high-resolution attention maps with features from the U-Net encoder, yielding per-pixel representations that are both fine-grained and robust [2]. Second, they identify a hierarchical semantic progression during the denoising process: representations shift from part-level abstractions at earlier timesteps to object-level abstractions at later stages [2]. Leveraging that progression, the method adaptively selects an optimal timestep for each pixel rather than applying one timestep across the entire image [1][2]. The paper reports that this dynamic, hierarchical timestep selection consistently outperforms existing zero-shot segmentation baselines in extensive experiments [2]. The research was posted on arXiv, an open-access repository that hosts preprints across physics, mathematics, computer science, and related fields [6]. As of November 2024, arXiv was receiving about 24,000 submissions per month and had surpassed two million total articles by the end of 2021 [6]. The paper appears with the standard arXiv tooling, including links to Bibliographic Explorer, Connected Papers, and Litmaps, which allow readers to traverse citation networks and discover related work [1][5]. arXiv also provides integration with code and data repositories through its Labs framework, a community-collaboration initiative that sets guidelines for third-party tools while requiring adherence to openness and user-data privacy values [4][5]. The submission itself is not peer-reviewed, consistent with arXiv’s role as a preprint server where papers are moderated but not formally evaluated before posting [6].