The shift from implicit neural representations (like NeRFs) to explicit primitive-based representations (like 3D Gaussian Splatting) has revolutionized real-time 3D scene rendering. By rasterizing millions of microscopic 3D ellipsoids instead of querying a heavy neural network for every ray step, graphics pipelines can achieve hundreds of frames per second.
However, this speed comes at a steep cost: a massive "memory wall" and severe computational bloat. As noted by Hanson et al. in Speedy-Splat and Jiang et al. in AnySplat, mapping image pixels directly to 3D space or optimizing complex scenes accumulates millions of individual primitives. This structural inefficiency inflates storage requirements and bogs down rendering hardware, presenting a major bottleneck for interactive or robotic applications (Zhou et al. in EA3D).
To solve this, researchers are developing techniques to identify and eliminate primitive redundancy. The core insight is simple: explicit scene representations drastically over-parameterize the world to make optimization easy, leaving behind a vast population of primitives that contribute virtually nothing to the final image.
The Root of Redundancy: Why Bloat Occurs
To understand why pruning works, we must first understand why 3D Gaussian Splatting creates so many unnecessary primitives. During optimization, standard algorithms aggressively clone, split, and densify Gaussians to capture intricate geometric details and view-dependent appearance.
This process creates millions of primitives that share highly similar geometric or appearance attributes (Navaneet et al. in CompGS). As Li et al. point out in HRGS, a massive percentage of these elements end up buried beneath surfaces, shrunk to microscopic scales, or rendered nearly transparent. They consume storage space and slow down tile-based rasterization without providing any proportional benefit to visual fidelity (Guo et al. in ContextGS).
In fact, this bloat is often an artifact of the optimization process itself rather than an inherent representation limit. Ding et al. demonstrate in AdamW-GS that simply decoupling the optimization steps reduces primitive redundancy by 50% from the start, proving that standard pipelines generate millions of "accidental" primitives due to coupled parameter updates.
Quantifying Importance: What Makes a Primitive Redundant?
If a scene is filled with millions of redundant Gaussians, how do we safely choose which ones to delete? In a comprehensive taxonomy of the field, Cong Zhang et al. in their survey paper outline several criteria used to calculate a primitive's significance:
- Opacity and Size: Primitives that are nearly transparent or too small to cover a pixel fraction can be safely dropped.
- Gradient Magnitude: Gaussians that receive consistently low positional gradients over many training steps are likely sitting in uninformative or empty space.
- Ray Contribution: Instead of analyzing a primitive in isolation, we can measure its global impact on the rendered views. Niemeyer et al. introduce a ray contribution-based pruning technique in RadSplat that tracks how much alpha-blending weight a Gaussian actually contributes across all training camera rays. If its cumulative contribution falls below a threshold, it is permanently culled.
[Camera Ray] ---> [Gaussian A (High Opacity)] ---> [Gaussian B (Hidden/Occluded)]
| |
High Contribution Zero Contribution
(KEEP PRIMITIVE) (PRUNE PRIMITIVE)
From Heuristics to Differentiable Pruning
Early pruning methods relied on hard, post-hoc thresholding rules applied after training. The cutting edge of this research direction, however, focuses on differentiable pruning mechanisms that allow the system to actively learn what to discard during the optimization process.
For example, Wu et al. introduce a framework in FastAvatar that leverages a Gumbel-Softmax distribution to model the discrete decision of keeping or dropping a primitive. Because the Gumbel-Softmax trick is continuous and differentiable, the network can backpropagate reconstruction errors directly into the pruning mask. This enables the model to automatically shed over 50% of its primitives while maintaining real-time rendering speeds exceeding 240 FPS.
Taking inspiration from neuromorphic computing, Weixing Zhang et al. propose Spiking GS, which integrates spiking neurons into the pipeline. The discrete, all-or-nothing firing behavior of spiking neurons acts as an aggressive, low-cost filter that suppresses low-opacity noise and strips away redundant parameters, yielding cleaner surface reconstructions with a fraction of the memory footprint.
Why Pruning Improves More Than Just Speed
Intuitively, removing parts of a model should degrade its performance. Why, then, does aggressive primitive pruning frequently maintain or even improve visual quality?
1. Regularization and Anti-Overfitting
In scenarios with limited data, such as sparse-view synthesis, dense clusters of primitives easily overfit to the exact training camera angles, creating floating artifacts that ruin novel views. As Shuai Zhang et al. show in TOGS and Han et al. show in Binocular-Guided 3DGS, randomly or strategically pruning non-essential spatial components acts as a powerful regularizer. It forces the remaining primitives to align tightly with the true underlying geometry of the scene.
2. Downstream Efficiency Gains
Pruning acts as a foundational step that unlocks massive downstream compression ratios. By eliminating uninformative primitives, the remaining "anchors" form a highly structured, clean spatial distribution. Yihang Chen et al. demonstrate in HAC that adaptive masking and pruning allow spatial context models to predict Gaussian attributes with extreme accuracy, enabling up to a 75× reduction in file size via entropy coding.
Furthermore, this newly cleared parameter space can be repurposed for other tasks. Xu et al. show in Mark3DGS that perception-aware pruning creates optimal capacity to embed robust, imperceptible intellectual property watermarks directly into the 3D scene geometry without degrading the final render.
Key Foundational Work
RadSplat: Radiance Field-Informed Gaussian Splatting for Robust Real- Time Rendering with 900+ FPS (2025) — contributed 2 concept elements to this direction.
Trick-GS: A Balanced Bag of Tricks for Efficient Gaussian Splatting (2025) — contributed 2 concept elements to this direction.
A Hierarchical Compression Technique for 3D Gaussian Splatting Compression (2025) — contributed 2 concept elements to this direction.