Frequency-Aware Gaussian Splatting Decomposition
Accepted to the International Conference on 3D Vision (3DV), 2026
Tel Aviv University, 2025
3D Gaussian Splatting (3D-GS) enables efficient novel view synthesis, but treats all frequencies uniformly, making it difficult to separate coarse structure from fine detail. Recent works have started to exploit frequency signals, but lack explicit frequency decomposition of the 3D representation itself. We propose a frequency-aware decomposition that organizes 3D Gaussians into groups corresponding to Laplacian-pyramid subbands of the input images. Each group is trained with spatial frequency regularization to confine it to its target frequency, while higher-frequency bands use signed residual colors to capture fine details that may be missed by lower-frequency reconstructions. A progressive coarse-to-fine training schedule stabilizes the decomposition. Our method achieves state-of-the-art reconstruction quality and rendering speed among all LOD-capable methods. In addition to improved interpretability, our method enables dynamic level-of-detail rendering, progressive streaming, foveated rendering, promptable 3D focus, and artistic filtering.
Our frequency-aware decomposition enables progressive Level of Detail (LOD) rendering, where low-frequency Gaussians capture coarse geometry and appearance, and higher levels progressively refine the image with finer details.
By grouping Gaussians according to Laplacian pyramid subbands, we isolate and visualize distinct frequency components of a scene, revealing how each level contributes to the final image structure.
Our model supports real-time foveated rendering by concentrating high-frequency details in the user’s gaze region, reducing rendering cost in the periphery without compromising perceptual quality.
Frequency-aware decomposition allows selective object editing based on text prompts by preserving only relevant high-frequency Gaussians, enabling multi-view-consistent object highlighting or suppression.
Manipulating frequency-specific Gaussian groups enables stylized effects such as X-ray or painterly rendering, while preserving 3D consistency and coherence across viewpoints.
Operating on only the low-frequency Gaussians enables faster geometric queries and interactions, such as collision detection or spatial search, while maintaining structural fidelity.
