EquiFusion - Kinematics-Agnostic Human Motion Prediction via Equivariant Latent Diffusion (ECCV 2026)
This is the official repository for the ECCV 2026 paper "EquiFusion: Kinematics-Agnostic Human Motion Prediction via Equivariant Latent Diffusion".
Authors: Cecilia Curreli, Florian Hofherr, Dominik Muhle, Abhishek Saroha, Riccardo Marin, Daniel Cremers
The model and training code for EquiFusion will be released here. Star or watch this repository to be notified when it lands. In the meantime, see the paper and the project page.
We address kinematics-agnostic (skeleton-agnostic) stochastic 3D human motion prediction: given a past motion, predict plausible futures for any skeleton, without hard-coding its kinematics. EquiFusion is a permutation-equivariant latent diffusion model that treats the skeleton's kinematic connectivity as an explicit input, so its internal computations are agnostic to joint ordering and graph structure.
This design enables cross-dataset / cross-skeleton generalization to unseen kinematics, and unlocks zero-shot directions such as motion prediction from partial or occluded observations and targeted limb generation — while remaining up to 75% more compact than kinematics-specific baselines, with faster training and inference.
This code is released under the BSD 2-Clause License. See LICENSE for details.
If you find this work useful, please cite:
@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}
}