Cross-Morphology Motion Transfer with Semantic Style Alignment
Overview
In this work, we present a neural framework for cross-morphology motion transfer that enables realistic animation retargeting across characters with diverse skeletal structures. By using morphology-agnostic control signals and morphology-specific latent codebooks, our method preserves both the semantic intent of the source motion and the natural style of the target character. We demonstrate accurate and scalable all-to-all retargeting across bipeds and quadrupeds, with optional user control for fine-grained motion editing.
Abstract
Transferring animations between characters with diverse skeletal structures is challenging. Traditional retargeting pipelines rely on fixed correspondences, canonical skeletons, or human-centric datasets, which can lead to artifacts when applied across heterogeneous morphologies. We introduce a framework for cross-morphology motion transfer with semantic style alignment that uses morphology-agnostic control signals (e.g., velocity, angular velocity, relative height) to align behaviors across species. Our method supports all-to-all retargeting: motions from any source can be mapped to any trained target while preserving target-specific style. For each target morphology, we train a Vector Quantized VAE and an autoregressive sequence model to construct a compact, morphology-specific codebook that captures stylistic priors. This modular design scales to new morphologies without retraining existing models and allows optional user control (e.g., phase, velocity scaling) for fine-grained alignment. Experiments across bipeds and quadrupeds demonstrate accurate, plausible, and style-faithful motion transfer, establishing a scalable approach to retargeting across arbitrary skeletal topologies.
Architecture
Figure 1: Overview of the two-stage training pipeline. (Left) Stage 1: Motion Generator. We train a skeleton-aware VQ-VAE to reconstruct full root-relative motion M and root trajectory R from sparse inputs Ms and skeletal offsets S. This stage learns a discrete, morphology-specific codebook C that captures the valid pose manifold. (Right) Stage 2: Autoregressive Code Predictor. The Motion Generator is frozen (indicated by lock icons) and processes the input motion to provide ground-truth code indices from codebook C. The predictor learns to forecast these indices autoregressively using a secondary, lightweight codebook C′, conditioned on extracted control signals. The interaction between the two stages is explicitly enforced by minimizing the cross-entropy loss (shown as the dashed line).
Video
Contributions
- A morphology-specific motion synthesis framework for all-to-all cross-morphology retargeting, which generates target motions from control signals of any source morphology without requiring joint- or phase-level correspondences. Semantic alignment is achieved through morphology-agnostic signals (e.g., velocity, angular velocity, relative height), with user-controllable cues (e.g., phase, velocity scaling) for intuitive style and pace adjustment.
- A two-stage training procedure combining VQ-VAE-based reconstruction with autoregressive sequence modeling. This design yields compact, morphology-specific codebooks that capture stylistic priors, enabling scalable extension to new characters without retraining existing ones.
- A comprehensive evaluation across heterogeneous morphologies (bipeds ↔ quadrupeds), demonstrating kinematic accuracy, physical plausibility, and faithful preservation of target-style dynamics.
BibTeX
@article{Mylordos:2026,
author = {Mylordos, Alexios and Ponton, Jose Luis and Pelechano, Nuria and Aristidou, Andreas},
journal = {IEEE Transactions on Visualization and Computer Graphics},
title = {Cross-Morphology Motion Transfer with Semantic Style Alignment},
year = {2026},
volume = {},
number = {},
pages = {1-14},
keywords = {Cross-morphology; motion transfer; retargeting; semantic alignment; stylistic alignment},
doi = {10.1109/TVCG.2026.3729292}
}
Acknowledgements
We would like to thank Theodoros Kyriakou for his assistance with the visualizations of our results. This project has received funding from the European Commission's Horizon Europe program under grant agreement 101178362, the European Union’s Horizon 2020 Research and Innovation Programme under Grant Agreement No 739578, the Government of the Republic of Cyprus, and the MCIN/AEI/10.13039/501100011033/FEDER “A way to make Europe”, UE (PID2021-122136OB-C21).
