ETA-MotionEfficient Topology-Agnostic Motion Diffusion
via Sparse Joint Learning

Seoyoung Hwang1,2Tae-Kyun Kim1

1 KAIST2 Samsung Electronics

WACV 2027

arXiv View mesh results Code Read abstract
Sparse joint skeletons of a quadruped and a bird used for training, followed by full rigged horse and eagle motion at inference.
ETA-Motion is trained on sparse joint subsets and generates motion for full target skeletons.
Examples of transfer from sparse skeletal observations to rigged characters.

Abstract

Topology-agnostic motion generation aims to synthesize motions for diverse characters with arbitrary skeletons. Unlike conventional motion generation, which assumes a fixed skeletal template, this setting must handle diverse joint configurations and inconsistent joint semantics across characters. Existing methods address this challenge through dense full-joint modeling, which incurs substantial computational cost and hinders learning transferable motion patterns across diverse skeletal topologies. We present ETA-Motion, a topology-agnostic motion generation framework based on sparse joint learning. Instead of learning motion from all joints, our method uses a compact subset of joints to reduce redundancy and trains the motion generator on the resulting sparse skeleton. To facilitate correspondence across diverse topologies, we introduce semantic joint embeddings that combine textual and visual cues, together with topology-aware training objectives for preserving skeletal structure. Although ETA-Motion is trained only on sparse skeletal observations, it supports full-skeleton motion generation during inference. Experimental results demonstrate that ETA-Motion outperforms existing topology-agnostic baselines while significantly improving efficiency and generalization to unseen motions and unseen topologies.

Method overview

A text-conditioned diffusion transformer learns motion from compact skeletal observations, with semantic cues that connect joints across different characters.

ETA-Motion architecture: textual T5 and visual DINOv2 joint embeddings condition a spatiotemporal diffusion transformer with sparse training joints, full inference skeletons, and a topology-aware loss.
ETA-Motion combines sparse joint learning, semantic joint embeddings, and a topology-aware training objective.
01

Select representative joints

Farthest point sampling retains a compact subset, starting from the root. Each selected joint reconnects to its nearest retained ancestor.

02

Connect joint semantics

T5 joint-name features and DINOv2 visual features provide complementary cues for correspondence across different skeletal topologies.

03

Generate the full motion

The target rest pose guides inference on the full skeleton. A bone-length objective encourages structural consistency during training.

Sparse joint selection

Training uses K joints instead of the full skeleton, reducing spatial attention from O(FJ2) to O(FK2). Inference uses the actual number of joints in the target skeleton.

A full quadruped skeleton, representative joints selected in green, and the reconnected sparse skeleton.

Mesh animation results

Generated joint positions are converted into mesh animations using inverse kinematics. Explore the motions across different characters below.

BibTeX

@inproceedings{hwang2027etamotion,
  title  = {{ETA-Motion}: Efficient Topology-Agnostic Motion
            Diffusion via Sparse Joint Learning},
  author = {Hwang, Seoyoung and Kim, Tae-Kyun},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV)},
  year = {2027}
}