AniMo: Species-Aware Model for Text-Driven Animal Motion Generation

Introduction

Text-to-motion generation has seen significant progress, yet most existing models remain "human-centric." Humans share a uniform skeletal structure, but animal morphology varies wildly—a "walk" for a crocodile is biomechanically distinct from that of a giraffe. Existing models often fail to capture these species-specific dynamics due to morphological diversity and a lack of high-quality, text-paired animal motion data.

AniMo addresses these challenges by introducing a species-aware architecture and a massive new dataset, AniMo4D, enabling the generation of realistic 3D motions for over 114 animal species.

Expand

Source, p. 1

Methodology

AniMo operates in two stages: Motion Tokenization and Text-to-Motion Generation.

Motion Tokenization

The model first compresses continuous 3D motion into discrete tokens using a Joint-Aware Spatiotemporal Encoder. This encoder utilizes a tree-structured attention mask rooted at the pelvis to respect physical skeletal connections.

A critical component is the Species-Aware Feature Modulation (SAFM). By incorporating species and attribute information (e.g., age, gender), SAFM dynamically scales motion features. This allows the architecture to adapt to varying scales, from the rapid movements of small animals to the heavy strides of large mammals.

Expand

Source, p. 3

Text-to-Motion Generation

In the second stage, a Base Transformer predicts motion tokens from text prompts using masked modeling, while a Residual Transformer refines high-frequency details. Both are conditioned on CLIP text embeddings that include species-specific data.

The AniMo4D Dataset

To train AniMo, the authors introduced AniMo4D, a large-scale dataset curated from Planet Zoo. It contains 78,149 motion sequences and 185,435 textual descriptions across 114 species. Unlike previous datasets, it includes attributes like age and gender, which are vital for realistic animation.

Key Results

AniMo significantly outperforms state-of-the-art baselines:

• Motion Quality: On the AniMo4D test set, AniMo achieved a Frechet Inception Distance (FID) of 0.029, vastly superior to the next best baseline (0.107).
• Text Alignment: The model reached a Top-1 R-Precision of 77.4%, demonstrating a strong ability to match generated motion to complex descriptions.
• Generalization: Testing on the out-of-distribution AnimalML3D dataset confirmed that AniMo's species-awareness generalizes beyond its training data.

Conclusion

AniMo represents a major step toward universal animal motion synthesis. By explicitly modeling morphology through SAFM and leveraging the AniMo4D dataset, the framework captures the nuances of animal behavior that human-centric models miss. Future work will look toward real-world motion capture integration and multi-animal interactions.

@inproceedings{wang2025animo,
  author = {Xuan Wang and Kai Ruan and Xing Zhang and Gaoang Wang},
  title = {{AniMo: Species-Aware Model for Text-Driven Animal Motion Generation}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2025},
  doi = {10.1109/cvpr52734.2025.00186},
}