FlowFeat:
Pixel-Dense Embedding of Motion Profiles


Nikita Araslanov
TU Munich / MCML
Anna Sonnweber
TU Munich
Daniel Cremers
TU Munich / MCML

✨ NeurIPS 2025 Spotlight ✨


Paper Supplemental Material Code (coming soon)
The gist: Distilling optical flow networks into pixel-level task-agnostic representations.

Input
DINOv2
V-JEPA
VideoMAE
FeatUp
FlowFeat
FlowFeat provides versatile pixel-level features. Using motion-driven embedding statistics, it achieves high spatial precision and temporal consistency. Shown: PCA visualization of feature maps vs. state-of-the-art encoders.

Video Object Segmentation

FlowFeat is a fine-grained feature representation, where dynamic objects are particularly prominent.

Comparison to FeatUp (Fu et al., 2024)

FlowFeat scales well with the input resolution – here, increased by a factor of 2.

Comparison to LoftUp (Huang et al., 2025)

FlowFeat fares well against supervised models such as LoftUp and can be trained in a completely unsupervised fashion.

Citation

@inproceedings{Araslanov:2025:FlowFeat,
  author = {Araslanov, Nikita and Sonnweber, Anna and Cremers, Daniel},
  title = {{FlowFeat}: Pixel-Dense Embedding of Motion Profiles},
  booktitle = {NeurIPS},
  year = {2025},
}