Cross-reenactment transferring motion and expression from one video to a different person's image. This demonstrates the model's ability to generalize across different identities.

3. Dynamic Novel Views (Rotation Only)

File: demo_videos/dynamic_novel_views_rotation.mp4 (1.7MB)

Dynamic novel view synthesis with camera rotation around the avatar. The camera orbits around the subject while maintaining a fixed distance, showing the 3D consistency of the reconstructed avatar.

4. Dynamic Novel Views (Rotation + Zoom) ⭐

File: demo_videos/dynamic_novel_views_rotation_zoom.mp4 (1.8MB)

This is the video with both rotation AND zoom!

Dynamic novel view synthesis combining:

Camera rotation: Horizontal and vertical rotation around the avatar
Camera zoom: Smooth zoom in/out effect (0.7x to 1.3x distance variation)

This demonstrates the full 3D reconstruction quality with both viewpoint changes and distance variations.

🎯 Features Demonstrated

Feature	Self	Cross	Rotation	Rotation+Zoom
Identity preservation	✅	✅	✅	✅
Motion transfer	✅	✅	✅	✅
Novel viewpoints	❌	❌	✅	✅
Distance variation	❌	❌	❌	✅

🚀 Technical Details

Model: GUAVA (286.98M parameters)
Input: Tracked video with SMPL-X and FLAME parameters
Output: 512x512 rendered frames at 30 FPS
Rendering Speed: ~6-9 iterations/second on single GPU

📖 Citation

If you find this work helpful, please cite:

bibtex

@article{GUAVA,
  title={GUAVA: Generalizable Upper Body 3D Gaussian Avatar},
  author={Zhang, Dongbin and Liu, Yunfei and Lin, Lijian and Zhu, Ye and Li, Yang and Qin, Minghan and Li, Yu and Wang, Haoqian},
  journal={arXiv preprint arXiv:2505.03351},
  year={2025}
}

🔗 Links

📝 License

MIT License

openhe/GUAVA-trans

GUAVA Demo Videos

📹 Demo Videos

1. Self-Reenactment

2. Cross-Reenactment