FLOWING 🌊: Implicit Neural Flows
for Structure-Preserving Morphing

Overview

Given source and target images \(I^0, I^1\) we extract landmark pairs \((p^0_i,p^1_i)\) with a feature extractor. We train a flow \(\phi\) such that \(\phi(p^0,t) = \phi(p^1,t-1)\), effectively mapping \(p^0\) to \(p^1\). At inference, to blend the images at time \(t\), we warp \(I^0\) forward by \(t\) units and \(I^1\) and backward by \(t-1\) units, then blend them together with methods such as linear blending or generative models.

Features

Our approach supports both images and 3D Gaussian data. Below, we show morphing between both face images and 3D Gaussian Splatting faces.

Proper warping before blending may be used for better feature transfer. On the example below, we morph two pictures of Albert Einstein, one real and a sketch. Notice how the sketch's traces are more present on our morphing (third column), compared to state-of-the-art generative morphing (diffmorpher, second column)

Below, we interpolate different views of the same building of the MegaDepth dataset, notice how the our method preserves the chimneys and center/rightmost antennaes (it unfortunately misses the leftmost antenna).

Paper

FLOWING 🌊: Implicit Neural Flows for Structure-Preserving Morphing

Arthur Bizzi, Matias Grynberg Portnoy, Vitor Pereira Matias, Daniel Perazzo, João Paulo Silva do Monte Lima, Luiz Velho, Nuno Gonçalves, João M. Pereira, Guilherme Schardong, Tiago Novello

description Paper preprint (PDF)

description Arxiv

insert_comment BibTeX

Citation

@inproceedings{bizzi2025flowing,
  title={{FLOWING}: Implicit Neural Flows for Structure-Preserving Morphing},
  author={Bizzi, Arthur and Grynberg Portnoy, Matias and Pereira Matias, Vitor and
      Perazzo, Daniel and Silva do Monte Lima, Jo\~{a}o Paulo and Velho, Luiz and
      Gon\c{c}alves, Nuno and Pereira, Jo\~{a}o and Schardong, Guilherme and Novello, Tiago},
  booktitle = {Advances in Neural Information Processing Systems},
  editor = {D. Belgrave and C. Zhang and H. Lin and R. Pascanu and P. Koniusz and
      M. Ghassemi and N. Chen},
  pages = {153041--153065},
  publisher = {Curran Associates, Inc.},
  year={2025},
}

Poster

Acknowledgements