README / README.md
ceccurr's picture
Organization card
a29adab verified
|
Raw History Blame Contribute Delete
2.74 kB
metadata
title: README
emoji: 馃暫
colorFrom: indigo
colorTo: green
sdk: static
pinned: false

EquiFusion

Kinematics-Agnostic Human Motion Prediction via Equivariant Latent Diffusion (ECCV 2026)

EquiFusion predicts diverse realistic futures from a given input past and supports any human skeleton or kinematics. Instead of converting or unifying skeletons of different formats or training a new instance for every skeleton type, we adopt an orthogonal solution: create a model that supports them all. We give the skeleton kinematic tree as input to the model together with the past motion, so a single model instance forecasts the future across different datasets and skeleton formats, also zero-shot on skeletons unseen during training and on observations with missing limbs or joints.

Paper 路 Code 路 Project page 路 Demo

Models

Model Training data Use
equifusion-cross-kinematics-amass-nymeria AMASS + Nymeria main model; zero-shot on other skeletons
equifusion-cross-kinematics-amass-nymeria-h36m AMASS + Nymeria + H36M trained on three datasets
equifusion-cross-kinematics-amass AMASS zero-shot on unseen skeletons
equifusion-cross-kinematics-amass-nymeria-3d AMASS + Nymeria 3D joint representation, arbitrary missing joints
equifusion-in-domain-amass AMASS same-dataset evaluation
equifusion-in-domain-nymeria Nymeria same-dataset evaluation
equifusion-in-domain-h36m H36M same-dataset evaluation

See the code repository for installation, inference, evaluation and training.

Citation

@inproceedings{curreli2026equifusion,
  title     = {EquiFusion: Kinematics-Agnostic Human Motion Prediction via Equivariant Latent Diffusion},
  author    = {Curreli, Cecilia and Hofherr, Florian and Muhle, Dominik and Saroha, Abhishek and Marin, Riccardo and Cremers, Daniel},
  booktitle = {European Conference on Computer Vision},
  organization = {Springer},
  year      = {2026}
}