Jonghoon Ahn.← Research portfolio
03 · Delivery · 3D Gaussian Splatting · Compression · Rate-Distortion

SplatStream Lab

Faster training still did not solve the delivery problem. This experiment asks how far a trained 3DGS scene can be reduced through pruning and quantization before held-out visual quality collapses.

gsplatCUDAMip-NeRF 360 BonsaiPSNR / SSIM / LPIPSFP16 / INT8zlib experimental entropy backend
Open GitHub repository

Why this came next

Reducing training cost does not help delivery if the finished representation is still too large.

The FastGS study showed that optimization time and model growth could be reduced, but the underlying delivery problem remained. A Gaussian scene is an explicit collection of many parameters. Even when it renders well, the representation can still be heavy enough that transfer cost becomes a first-class systems problem.

That shifted the research question from “how quickly can I train it?” to “how much information does the final scene actually need to retain?”

Research question

How much can a trained 3DGS scene be reduced through importance pruning and numerical precision changes before held-out image quality degrades sharply? I used a simple view-independent importance heuristic based on opacity and geometric mean scale, then reduced precision and applied a generic entropy-coding stage. The heuristic is intentionally presented as an experimental baseline, not a published compression algorithm.

Baseline payload185.43 MiB

Entropy-coded full-scene representation used as the comparison baseline.

Selected operating point70.04 MiB

25% prune + FP16. A 2.65× reduction from baseline.

Held-out quality27.62 dB

PSNR with 0.9007 SSIM and 0.1659 LPIPS.

Rate-distortion results

ConfigurationPayloadCompressionPSNR ↑SSIM ↑LPIPS ↓
Baseline185.43 MiB1.00×29.77050.92760.1508
25% prune + FP1670.04 MiB2.65×27.62340.90070.1659
50% prune + FP1646.78 MiB3.96×23.42290.81780.2364
50% prune + INT820.24 MiB9.16×17.23070.56530.5563
75% prune + INT810.31 MiB17.98×17.14990.56170.5638
Selected operating point: 2.65× smaller with useful retained quality.

The 25% prune + FP16 configuration reduced the experimental entropy-coded scene from 185.43 MiB to 70.04 MiB while retaining 27.62 dB PSNR, 0.9007 SSIM, and 0.1659 LPIPS.

What failed, and why that mattered

INT8 was not a free compression win.

The aggressive INT8 configurations produced much smaller payloads, but image quality dropped sharply. PSNR fell to roughly 17 dB and LPIPS increased substantially. I treat that as a quality cliff rather than hiding the configuration.

This negative result is useful because it marks the limit of a simple pruning-plus-quantization strategy. If the goal is to push substantially below the selected FP16 operating point, a better codec would likely need more structured parameter coding, learned transforms, entropy modeling, or representation-aware quantization instead of applying a uniform low-precision treatment.

Why this is a media systems question

A representation becomes streamable only when rate and distortion are considered together.

The smallest file is not necessarily the best result, and the highest-quality scene is not necessarily practical to deliver. The useful operating point lies between them. That is why this study reports model size beside PSNR, SSIM, and LPIPS instead of treating compression as a single-number objective.

The experiment does not claim a production streaming codec. It establishes a measured rate-distortion baseline that can support the next generation of encoding experiments.

What this led to next

Interpretation boundary

This is a one-scene rate-distortion study. zlib is used only as a generic experimental entropy backend and is not presented as a production 3DGS codec. Network transfer estimates are illustrative. The experiment does not claim Netflix-scale streaming infrastructure, production codec performance, or rendering acceleration.