What Moves?
Related Work

From the Authors

DisMo NeurIPS 2025 Semantic motion representations disentangled from static content for open-world motion transfer.

Related Work

Segment to Focus arXiv 2026 Mask-guided latent action learning focuses reconstruction on agent motion instead of visual distractors. Motion4Motion SIGGRAPH 2026 Training-free, skeleton-free motion transfer across subjects using dense motion flow.

What Moves?

Localized Motion Representations for Compositional Scene Control

Frank Fundel*, Malek Ben Alaya*, Thomas Ressler-Antal*, Stefan Andreas Baumann, Björn Ommer

* equal contribution

CompVis @ LMU Munich, MCML

Demo arXiv Code Weights

TL;DR: Real-world scenes contain multiple entities moving at once, but most video representations either encode motion globally or localize it by cropping away the context needed to interpret it. We introduce promptable localized motion representations: the encoder sees the full video, while masks query persistent region-specific motion tokens. These tokens support object-level motion transfer, compositional scene control, and localized action classification.

Contextualized Motion Representations

Steering a Video Model

Context and Locality Matters

Localized Action Classification

A2D, kNN over queried actor motion embeddings.

DINOv2 RGB 17.4
V-JEPA2 RGB 23.8
DisMo RGB 33.9
Ours 42.0

Controllable Motion Transfer

Motion selectivity: in-region fidelity minus leakage.

ATI 0.200
WanMove 0.183
DisMo 0.138
Ours 0.230

Global Motion Classification

With a full-frame query, the same encoder becomes a global motion representation and remains competitive with dedicated video and motion models.

Model SSv2 Jester Diving48 ARID IARD
DINOv2 10.3 23.2 9.4 14.5 76.1
VideoMAE 7.1 20.1 - 17.3 73.4
V-JEPA2 22.2 40.8 11.1 28.0 87.6
SemanticMoments DINO 11.4 37.3 9.9 22.0 62.7
SemanticMoments V-JEPA2 31.5 52.2 12.7 38.3 92.5
DisMo 24.6 69.8 20.9 55.3 92.0
Ours 26.5 72.8 17.1 55.4 93.0

Qualitative Examples

Source A Source B Generation
Three starting scenes with the selected subjects outlined by masks
Three starting scenes with the selected subjects outlined by masks
Three starting scenes with the selected subjects outlined by masks
Three starting scenes with the selected subjects outlined by masks
Comparisons Open the gallery

Citation

BibTeX

@inproceedings{fundel2026whatmoves,
  title     = {What Moves? Localized Motion Representations for Compositional Scene Control},
  author    = {Fundel, Frank and Ben Alaya, Malek and Ressler-Antal, Thomas and Baumann, Stefan Andreas and Ommer, Bjorn},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}