💻 computer science

QIRF Quantum-Inspired Non-Orthogonal Function-Space Compression for 3D Gaussian Splatting

QIRF introduces a quantum-inspired, non-orthogonal function-space compression method for 3D Gaussian Splatting that leverages generalized eigendecomposition on an overlap matrix to identify and select representative primitives, achieving significant reductions in storage and rendering costs while maintaining or slightly improving reconstruction quality and speed.

Original authors: Shizeng Jiang, Hao Zhang, Xuerui Ma, Ying Hu, Tao Zhang

Published 2026-07-21

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shizeng Jiang, Hao Zhang, Xuerui Ma, Ying Hu, Tao Zhang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to build a perfect, photorealistic model of a city using millions of tiny, glowing, 3D paint blobs. In the world of computer graphics, this is exactly what a technology called 3D Gaussian Splatting does. It uses these "blobs" (mathematically known as anisotropic Gaussian primitives) to represent every corner, texture, and shadow of a scene. The result is stunning: you can walk around the model and see it from any angle in real-time, just like a video game. However, there's a catch. To make the picture look sharp and detailed, the computer often needs to use millions of these blobs. This creates a massive file that is hard to store, hard to send over the internet, and heavy for your computer's graphics card to process.

For a long time, the standard way to shrink these files was to just delete the "less important" blobs or to squish their numbers down to take up less space. But this paper suggests that approach is a bit like trying to organize a messy room by only throwing away the smallest items, while ignoring the fact that you have ten identical red chairs piled up in the corner. The authors realized that many of these glowing blobs overlap so much that they are essentially doing the exact same job. Instead of looking at each blob individually, they decided to look at the group of overlapping blobs as a single, tangled unit and figure out how to untangle it.

Enter QIRF (Quantum-Inspired Non-Orthogonal Function-Space Compression), a new method that treats these overlapping blobs like a complex dance troupe rather than a crowd of strangers. The researchers, led by Shizeng Jiang and Hao Zhang, noticed that when many blobs overlap, they create a kind of "functional redundancy"—they are all trying to paint the same part of the picture. To solve this, they borrowed a clever idea from quantum chemistry (the study of how atoms and electrons interact). In that field, scientists use math to figure out which atomic orbitals are truly unique and which are just repeating the same information.

QIRF applies this same logic to 3D scenes. Instead of asking, "Is this single blob important?" it asks, "If we look at this whole cluster of overlapping blobs, which few of them are actually doing the unique work?" It builds a mathematical map of how much the blobs overlap and how much they contribute to the final image. Then, using a special type of math called generalized eigendecomposition, it identifies the "dominant" blobs that carry the most important information and discards the ones that are just copying their neighbors.

But there's a risk: if you just delete the "copycat" blobs, you might accidentally erase the tiny, delicate details, like the thin spokes of a bicycle wheel or the edge of a leaf. To prevent this, QIRF uses a "detail-aware safeguard." It keeps a small safety net of extra blobs specifically for these tricky, high-frequency details, ensuring that the final image doesn't look blurry or lose its sharp edges.

The results are quite impressive. When the team tested QIRF on 13 different complex scenes (including famous datasets like Mip-NeRF 360 and Tanks&Temples), they found that it could remove an average of 71.7% of the Gaussian blobs. This shrank the raw file size by about 3.54 times (meaning the files are less than a third of their original size). Surprisingly, the image quality didn't just stay the same; it actually got slightly better, with an average improvement of 0.10 dB in image clarity (PSNR). Furthermore, because there were fewer blobs for the computer to draw, the scenes rendered 34.3% faster on average.

The paper suggests that by treating these 3D scenes as a "non-orthogonal function space" (a fancy way of saying a space where the building blocks overlap and aren't perfectly independent), we can find a much more efficient way to store and display 3D worlds. While the authors note that their method focuses on removing the redundant blobs and doesn't yet combine with other compression tricks like shrinking the numbers inside the blobs, their findings suggest that looking at the "group dynamics" of these 3D primitives is a powerful, under-explored way to make virtual reality and 3D graphics much lighter and faster.

Technical Summary: QIRF - Quantum-Inspired Non-Orthogonal Function-Space Compression for 3D Gaussian Splatting

Problem Statement

3D Gaussian Splatting (3DGS) achieves high-quality real-time rendering by representing scenes with explicit anisotropic Gaussian primitives. However, the adaptive densification process often results in millions of primitives for complex scenes, leading to substantial storage overhead (hundreds of megabytes to gigabytes in raw PLY format) and increased rendering costs.

Existing compression methods primarily address this through:

Primitive-wise pruning: Assigning independent importance scores based on opacity, visibility, or gradients.
Attribute compression: Quantization and entropy coding of Gaussian parameters.

The paper identifies a critical gap: these methods treat Gaussians as independent entities and fail to explicitly model functional redundancy caused by strongly overlapping, non-orthogonal Gaussian basis functions. Two Gaussians may both have high individual importance scores yet span nearly the same local scene function, representing redundant information in the underlying radiance field.

Methodology

The authors propose QIRF (Quantum-Inspired Non-Orthogonal Function-Space Compression), which reframes scene compression as a local non-orthogonal basis-reduction problem. Inspired by non-orthogonal basis analysis in quantum chemistry (specifically the Roothaan-Hall formulation and natural orbital analysis), the method operates as follows:

Adaptive Partitioning:
The scene is divided into local voxel blocks. The voxel size is adaptively selected based on the sparsity of the initial partition to ensure stable local subspace estimation.
Local Function-Space Modeling:
Within each block, neighboring Gaussians are treated as overlapping non-orthogonal basis functions. Two matrices are constructed:
- Gaussian Overlap Matrix ( $S$ ): An analytic matrix measuring the functional similarity (spatial overlap) between neighboring Gaussian bases.
- Radiance-Response Density Matrix ( $P$ ): A matrix aggregating rendering and optimization responses. It combines rendering contributions (visibility, composited contribution, etc.) and optimization gradients (position, scale, opacity, appearance) using an exponential moving average.
Generalized Natural Scene Orbital (GNSO) Analysis:
QIRF solves a generalized eigenvalue problem:
$P u_k = \lambda_k S u_k$
Unlike standard PCA which assumes an orthogonal basis, this formulation explicitly incorporates the overlap matrix $S$ as the metric of the local basis space. The eigenvectors ( $u_k$ ) define the Generalized Natural Scene Orbitals, and the eigenvalues ( $\lambda_k$ ) represent the occupation strength of each mode.
- Subspace Selection: The dominant local function subspace is identified by retaining the smallest number of modes that preserve a cumulative energy threshold (e.g., 95%).
- Primitive Selection: Representative Gaussians are selected based on their participation scores in the retained modes.
Detail-Aware Safeguarding:
To prevent the loss of high-frequency details (thin structures, occlusion boundaries) during aggressive pruning, a safeguarding step is applied. Inactive Gaussians are scored based on gradient response, visibility, scale, and opacity. A small fraction (e.g., top 5%) of high-scoring inactive Gaussians are retained alongside the GNSO-selected set.
Post-Pruning Optimization:
The selected primitives undergo fine-tuning with densification disabled. An additional reconstruction loss is introduced to compensate for quality degradation, ensuring a fixed-cardinality representation.

Key Contributions

Formulation: The paper formulates compact 3DGS as a local non-orthogonal basis-reduction problem, explicitly modeling functional redundancy among overlapping primitives.
Algorithm: It introduces a subspace-aware selection method using generalized eigendecomposition of an overlap matrix and a radiance-response density matrix.
Strategy: It develops a detail-aware safeguarding mechanism to preserve visually important structures under aggressive compression.
Performance: Experiments demonstrate significant reductions in storage and rendering costs while maintaining or slightly improving reconstruction quality.

Experimental Results

Evaluated on 13 scenes across Mip-NeRF 360, Tanks&Temples, and Deep Blending:

Compression Ratio: QIRF reduces the Gaussian count and raw PLY storage by an average of 71.7%, achieving approximately 3.54× compression.
Quality: It maintains reconstruction quality comparable to standard 3DGS, with a marginal average PSNR improvement of 0.10 dB.
Rendering Speed: By reducing the number of primitives, QIRF improves average rendering speed by 34.3% over vanilla 3DGS (though it is noted to be slower than highly optimized rendering-specific methods like MaskGaussian).
Storage: It achieves the smallest raw PLY size among compared methods (e.g., 224.96 MB on Mip-NeRF 360 vs. 287.38 MB for MaskGaussian).
Ablation: The study confirms that the generalized eigenspace formulation outperforms ordinary eigenspace selection, and the detail-aware safeguarding significantly improves SSIM and LPIPS metrics with minimal storage overhead.

Significance and Claims

The paper claims that non-orthogonal function-space redundancy is a significant yet underexplored source of representational redundancy in explicit Gaussian radiance fields. By shifting the perspective from independent primitive scoring to a function-space reduction problem, QIRF offers a principled way to remove redundant basis functions before attribute optimization.

The authors position QIRF as complementary to existing pruning and attribute-compression methods. They acknowledge limitations, noting that the current framework focuses on primitive-level redundancy and does not yet integrate attribute quantization or specialized sparse rasterization kernels, which are left for future work to further enhance efficiency.