Our group's new paper has been accepted to NeurIPS 2025 as a Spotlight. The work tackles a key limitation of Vision Transformers by replacing fixed patch tokens with an end-to-end differentiable tokenizer that adapts to image content at pixel-level granularity. Crucially, the approach is backward-compatible, enabling seamless retrofitting of existing pretrained models. Using hierarchical model selection guided by information criteria, the method delivers competitive results on both image-level classification and dense-prediction tasks. It also enables out-of-the-box raster-to-vector conversion, opening doors to more efficient and semantically aware vision pipelines.
See you at NeurIPS 2025!