Paper presented at NeurIPS 2025

The paper entitled “FLOWING: Implicit Neural Flows for Structure-Preserving Morphing“, co-authored by Prof. Nuno and Dr. Guilherme, was presented last December 3 at EurIPS 2025 and December 4 at NeurIPS 2025.

This year, NeurIPS 2025, one of the premier conferences in machine learning and artificial intelligence, continued to be extremely competitive. The Main Program received 21,575 valid paper submissions and accepted 5,290 papers—an acceptance rate of 24.5% on par with previous years despite the massive increase in submissions.

Morphing is a long-standing challenge in vision and graphics, often requiring feature-aligned warping and heavy regularization when using MLPs as implicit neural representations. These approaches are computationally costly and frequently unstable. FLOWING reframes morphing as a differential vector flow, embedding structural flow properties directly into the network. This naturally enforces continuity, invertibility, and temporal coherence. The result is a stable, principled method for structure-preserving morphing of both images and 3D Gaussian Splatting (3DGS).

Guilherme and Nuno would like to thank Fundação de Ciência e Tecnologia (FCT) projects UIDB/00048/20202 and UIDP/00048/2020 for partially funding this work. Guilherme would also like to thank FCT project 2024.07681.IACDC3 for partially funding this work. João Paulo would like to thank Fundação Carlos Chagas Filho de Amparo à Pesquisa do Estado do Rio de Janeiro (FAPERJ) grant SEI-260003/012808/2024 for funding this work. Vitor and Daniel gratefully acknowledge support from CAPES, grants 88887.842584/2023-00 and 88887.832821/2023-00, respectively for supporting this research. We also thank Google for funding this research.