Softmax-GS: Generalized Gaussians Learning When to Blend or Bound

Chen Ziwen, Peng Wang, Hao Tan, Zexiang Xu, Li Fuxin

TL;DR

Softmax-GS introduces a learnable, softmax-based competition mechanism for overlapping 3D Gaussians, resolving the long-standing "popping artifacts" and blurry boundaries inherent in standard 3D Gaussian Splatting (3D GS). By allowing Gaussians to adaptively blend colors or form crisp boundaries while maintaining order invariance, it achieves state-of-the-art novel view synthesis with significantly higher parameter efficiency.

What Problem Does This Paper Address?

3D Gaussian Splatting has revolutionized novel view synthesis due to its blazing-fast training and rendering speeds. However, its efficiency hinges on a mathematically convenient but physically unrealistic assumption: that 3D Gaussians do not overlap in space.

When the camera rotates, the global depth-sorting order of intersecting Gaussians abruptly changes. This violation of the no-overlap assumption causes distracting visual "popping" artifacts and view inconsistencies. Furthermore, standard Gaussian kernels have inherently diffuse, blurry tails, making them highly inefficient at representing sharp object boundaries without stacking millions of tiny primitives.

Figure 1. Comparison between different versions of 3D GS and Softmax-GS under slight left and right rotation. Vanilla 3D GS suffers from diffuse boundary and view inconsistency (“popping effect”) due to the no-overlap assumption. Only resorting[25] cannot fix view inconsistency for Gaussians on the same surface. Softmax-GS provides both boundary sharpness control and view- point consistency between overlapping Gaussians. 360° rotation videos are available on the project page.

Methods

Softmax-GS relaxes the no-overlap assumption by introducing a softmax competition between overlapping Gaussians directly within the volume rendering formulation.

  • Softmax Competition: Instead of standard alpha blending, the contribution of overlapping Gaussians is modulated by a softmax function over their exponents, controlled by a learnable sharpness parameter $\beta$. A high $\beta$ leads to a "winner-take-all" crisp boundary, while a low $\beta$ allows smooth color blending.
  • Order Invariance & Transmittance Maintenance: Simply applying a softmax along a ray breaks standard rasterization assumptions and introduces sequential bias. The authors derive a mathematical correction that ensures the rendering output is invariant to the sorting order of overlapping Gaussians and preserves the total scene transmittance.
  • Controllable Sharpness: To sharpen individual primitive edges, they integrate the Generalized Exponential Function (GEF) directly into the CUDA kernel, replacing the standard Gaussian squared distance with a learnable power parameter ($\alpha$).
  • Efficient Caching: To maintain linear time complexity during the backward pass, they cache intermediate values for only the closest $K=128$ Gaussians per image patch, avoiding full numerical integration.
Figure 2. Softmax-GS provides flexible boundary sharpness control and introduces a viewpoint-consistent, softmax-based color-merging mechanism for overlapping Gaussians, enabling both smooth color blending and a winner-take-all behavior.

Datasets

The authors evaluated their method on both synthetic and real-world benchmarks:

  • Synthetic Geometry Patterns: Custom 2D datasets (circle4, rect4, oval4) designed to explicitly test sharp edge reconstruction and overlapping behavior under tight primitive constraints.
  • Real-World Benchmarks: Canonical datasets including Mip-NeRF360 (7 scenes), Tanks&Temples (train and truck), and DeepBlending (drjohnson and playroom).

Key Insights and Ideas

  • Interaction Matters More Than Primitives: The core "aha moment" is that boundary sharpness cannot be solved purely by modifying individual primitives (like changing Gaussians to cylinders or boxes). True edge fidelity requires modeling the interaction and competition between adjacent primitives in 3D space.
  • The Blend-or-Bound Spectrum: By parameterizing the softmax harshness, the network doesn't have to choose between hard surfaces or volumetric fog globally; it learns exactly where to blend smoothly (e.g., color gradients) and where to bound sharply (e.g., object silhouettes).

Key Results

  • Extreme Sparsity Robustness: When restricted to a stringent budget of just 5% of the standard number of Gaussians, Softmax-GS outperforms vanilla 3D GS by over 1 dB PSNR across all real-world datasets.
  • Halving Parameter Count: The lightweight variant, Softmax-GSlight, achieves rendering quality on par with vanilla 3D GS while utilizing roughly half the number of Gaussians.
  • State-of-the-Art Fidelity: When integrated with advanced optimization frameworks like GS-MCMC, Softmax-GS achieves top-tier reconstruction fidelity (+0.3 dB PSNR improvement over standard 3D GS) while maintaining a rendering speed close to the original framework (~80% FPS).
Figure 5. Simple geometry fitting with 4 Gaussians using 3D GS, Softmax-GS, and two variants ablating softmax competition or boundary sharpness. Only one component yields suboptimal results, while Softmax-GS achieves the best of both. Optimization process videos are available on the project page.

Limitations and Future Work

To maintain the linear time complexity required for real-time rendering during the backward pass, the softmax competition is restricted to the closest $K=128$ Gaussians per patch. While this empirical threshold covers 70% to 93% of the Gaussians in the evaluated scenes, it could become a bottleneck in extremely complex environments with massive depth complexity (e.g., dense foliage or hair). The authors suggest that future work could explore adaptive selection strategies for $K$ to handle these highly complex scenarios.

Summary updated

Abstract

3D Gaussian Splatting (3D GS) is a popular representation for novel view synthesis due to its fast training and rendering speed. However, its efficiency relies on the key assumption that Gaussians do not overlap in the 3D space, which leads to noticeable artifacts and view inconsistencies. In addition, the inherently diffuse boundaries of Gaussians hinder accurate reconstruction of sharp object edges. We propose Softmax-GS, a unified solution that addresses both the view-inconsistencies and the diffuse-boundary problem by enforcing a softmax-based competition in overlapping regions between two Gaussians, with learnable parameters controlling the sharpness of the competition. This enables a continuous spectrum from smooth color blending to crisp, well-defined boundaries. Our algorithm also preserves order invariance among competing Gaussians and maintains the final transmittance. Extensive experiments on synthetic geometries demonstrate the effectiveness of each component of Softmax-GS, and evaluations on canonical real-world benchmarks show that it achieves state-of-the-art performance, improving both reconstruction quality and parameter efficiency. Project page: http: //arthurhero.github.io/projects/smgs/
BibTeX
@inproceedings{ziwen2026softmax,
  author = {Chen Ziwen and Peng Wang and Hao Tan and Zexiang Xu and Li Fuxin},
  title = {{Softmax-GS: Generalized Gaussians Learning When to Blend or Bound}},
  booktitle = {IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  year = {2026},
  url = {https://openreview.net/forum?id=FR8Ck6i2zZ},
}
Figure