TED-4DGS: Temporally Activated and Embedding-based Deformation for 4DGS Compression

TL;DR

TED-4DGS addresses the challenge of efficiently compressing dynamic 3D scenes by integrating temporally activated, embedding-based deformation into a sparse anchor-based 4D Gaussian Splatting framework. It couples per-anchor temporal features with a shared global deformation bank, explicit temporal activation parameters for occlusion handling, and an INR-based hyperprior with channel-wise autoregression for entropy coding. Empirical results on Neu3D and HyperNeRF demonstrate state-of-the-art rate-distortion performance and substantial bitrate reductions while preserving rendering quality. The approach advances RD-optimized dynamic 4DGS representations by tackling temporal visibility, deformation signaling, and compact attribute coding in a unified framework.

Abstract

Building on the success of 3D Gaussian Splatting (3DGS) in static 3D scene representation, its extension to dynamic scenes, commonly referred to as 4DGS or dynamic 3DGS, has attracted increasing attention. However, designing more compact and efficient deformation schemes together with rate-distortion-optimized compression strategies for dynamic 3DGS representations remains an underexplored area. Prior methods either rely on space-time 4DGS with overspecified, short-lived Gaussian primitives or on canonical 3DGS with deformation that lacks explicit temporal control. To address this, we present TED-4DGS, a temporally activated and embedding-based deformation scheme for rate-distortion-optimized 4DGS compression that unifies the strengths of both families. TED-4DGS is built on a sparse anchor-based 3DGS representation. Each canonical anchor is assigned learnable temporal-activation parameters to specify its appearance and disappearance transitions over time, while a lightweight per-anchor temporal embedding queries a shared deformation bank to produce anchor-specific deformation. For rate-distortion compression, we incorporate an implicit neural representation (INR)-based hyperprior to model anchor attribute distributions, along with a channel-wise autoregressive model to capture intra-anchor correlations. With these novel elements, our scheme achieves state-of-the-art rate-distortion performance on several real-world datasets. To the best of our knowledge, this work represents one of the first attempts to pursue a rate-distortion-optimized compression framework for dynamic 3DGS representations.

Figures (7)

• Figure 1: Overview of TED-4DGS.Left: Qualitative comparison on a banana scene. Our TED-4DGS reconstructs the scene with superior rendering quality compared to ADC-GS. It achieves a 26% file size reduction while closely matching the ground-truth view. Centre: Temporal duration map. Static background regions reuse long-duration Gaussian primitives, whereas occluded parts of the hand and banana are represented by short-duration primitives, demonstrating the effectiveness of temporal activation. Right-top: Rate-distortion comparison on the HyperNeRF park2021hypernerfhigherdimensionalrepresentationtopologically benchmark. Our TED-4DGS attains higher PSNR with smaller file sizes than prior methods. Right-bottom:Illustration of the learnable temporal-activation function, which activates a Gaussian primitive from its appearance ($a_s$) to disappearance ($a_f$).
• Figure 2: System overview of our TED-4DGS framework.
• Figure 3: Rate-distortion comparison of our TED-4DGS, Light4GS liu2025light4gslightweightcompact4d and ADC-GS huang2025adcgsanchordrivendeformablecompressed.
• Figure 4: Subjective quality comparisons.
• Figure 5: Rate-distortion comparisons on (a) deformation field variants and (b) compression variants.