How to Reduce Gaussian Splatting File Size: Compression Guide
Updated Mar 2026
A single 3D Gaussian Splatting scene can produce PLY files of 200 MB to 2 GB — impractical for sharing, web hosting, or mobile viewing. But the same scene can be compressed to 10-50 MB without visible quality loss using the right format and optimization pipeline. This guide covers every technique for reducing 3DGS file size: format-based compression (SPZ, SOG), Gaussian count reduction, floater removal, and SH degree selection. Each technique includes real compression ratios from actual scenes, not theoretical numbers.
Tools used in this guide
Step-by-Step Guide
- 1
Step 1 — Inspect Your File First
Before optimizing, understand what you are working with. Use our 3DGS File Inspector (polyvia3d.com/splat-inspector) to check: Gaussian count (determines base size), SH degree (degree 3 = 48 SH coefficients = ~236 bytes/Gaussian; degree 0 = 3 coefficients = ~62 bytes/Gaussian), format (PLY, SPLAT, SPZ, KSplat, SOG), and estimated GPU memory. This tells you where the size is coming from and which optimization will have the biggest impact.
- 2
Step 2 — Remove Floaters (Free Size Reduction)
Every 3DGS scan contains floating artifacts — stray Gaussians that contribute nothing to visual quality but add to file size. Our automatic Floater Removal tool (polyvia3d.com/splat-cleanup) uses Statistical Outlier Removal to identify and delete these. In testing, floater removal typically reduces Gaussian count by 3-8% with no visible quality loss. On a 1M Gaussian scene, removing 50K floaters saves 12 MB in raw PLY. This is free optimization — you are removing noise, not useful data.
- 3
Step 3 — Choose the Right Compression Format
This is where the biggest savings come from. Comparison for a 1M Gaussian scene with SH degree 3 (raw PLY = 236 MB): SPLAT format: 32 MB (86% reduction). Drops all SH data — view-dependent colors gone, flat appearance. Fast to load, broadest viewer support. SPZ format: 24 MB (90% reduction). Preserves all SH data via quantization + gzip. Visual quality nearly identical to PLY. Recommended for most use cases. SOG format: 12 MB (95% reduction). Morton ordering + WebP textures + codebook quantization. Best compression, fastest GPU decode. Requires viewers that support SOG (Spark.js, PlayCanvas). Use SPZ if you need wide compatibility, SOG if you need maximum compression.
- 4
Step 4 — Reduce SH Degree (If Quality Allows)
Spherical harmonics control view-dependent color — how the scene looks from different angles. SH degree 3 (48 coefficients per Gaussian) provides the best quality but uses 192 bytes per Gaussian just for SH data. SH degree 1 (12 coefficients) reduces this to 48 bytes — a 75% reduction in SH storage. SH degree 0 (3 coefficients) uses only 12 bytes — flat colors, no view-dependent effects. If your scene is viewed from a fixed camera (e.g., a product showcase), you can retrain at lower SH degree and save significant file size. For most scenes viewed from multiple angles, keep SH degree 2 or 3.
- 5
Step 5 — Reduce Gaussian Count
If compression alone is not enough, reduce the number of Gaussians. Our LOD tool (polyvia3d.com/splat-lod) ranks every Gaussian by visual importance (opacity times scale) and keeps only the top N%. At 75%, most scenes look nearly identical. At 50%, you see slight softening in fine details. At 25%, the scene is visibly simplified but the overall shape is preserved. For mobile delivery, a 50% reduction + SPZ compression can bring a 1M Gaussian scene from 236 MB down to 12 MB. Combine with floater removal for maximum effect.
- 6
Step 6 — Verify Quality After Optimization
After each optimization step, verify the result has not degraded unacceptably. Use our Quality Compare tool (polyvia3d.com/splat-compare) to load the original and optimized files side by side with synchronized cameras. Orbit through the scene and look for: color banding (artifact of heavy quantization), missing thin structures (first casualty of Gaussian count reduction), and flat spots where view-dependent color should be visible (sign that SH data was lost). If quality is unacceptable, back off the most aggressive optimization and try a different combination.