Building on the success of 3D Gaussian Splatting (3DGS), representing dynamic scenes (4DGS) has become a key research focus. However, achieving efficient deformation modeling while simultaneously optimizing for compression (Rate-Distortion Optimization) remains a significant challenge. Prior methods often struggle with either overspecified, short-lived primitives or irregular deformations that lack explicit temporal control.
To address this, we present TED-4DGS, a novel framework that unifies Temporally Activated mechanisms with Embedding-based Deformation for high-efficiency 4DGS compression. Our approach introduces learnable temporal activation parameters for each anchor to explicitly model appearance and disappearance, effectively handling occlusion and disocclusion. Furthermore, we incorporate an INR-based hyperprior and a channel-wise autoregressive model to achieve compact attribute compression. Experimental results demonstrate that TED-4DGS achieves state-of-the-art rate-distortion performance on real-world datasets, significantly reducing file sizes while maintaining superior rendering quality compared to existing methods.
Our work, TED-4DGS, introduces a rate-distortion-optimized compression framework for dynamic 3DGS. We design a per-anchor embedding-based deformation network (a) that leverages temporal features to query a shared global deformation bank to effectively capture anchor-specific deformation. To promote stable deformation and improved visibility modeling, we extend static 3D anchors into 4D by introducing temporal activation parameters. Finally, we incorporate an INR-based hyperprior and a channel-wise autoregressive model (b) for efficient attribute coding.
Our TED-4DGS consistently outperforms the competing methods across both datasets(HyperNeRF and Neu3D), achieving the state-of-the-art RD performance.
@misc{ho2025ted4dgstemporallyactivatedembeddingbased,
title={TED-4DGS: Temporally Activated and Embedding-based Deformation for 4DGS Compression},
author={Cheng-Yuan Ho and He-Bi Yang and Jui-Chiu Chiang and Yu-Lun Liu and Wen-Hsiao Peng},
year={2025},
eprint={2512.05446},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2512.05446},
}
This work is supported by MediaTek Advanced Research Center and National Science and Technology Council (NSTC), Taiwan, under Grants 113-2634-F-A49-007-, 112-2221-E-A49-092-MY3, and 114-2221-E-A49-035-MY3. We thank National Center for High-performance Computing (NCHC) for providing computational and storage resources.