Research map / Point clouds / Point cloud classification
HybridGS: High-Efficiency Gaussian Splatting Data Compression using Dual-Channel Sparse Representation and Point Cloud Encoder
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes HybridGS, which generates explicit compact 3DGS via dual-channel sparsification (latent low-dimensional color/rotation features with quantization, and integer-position primitive pruning/uniqueness) and then compresses it with GPCC, achieving comparable rate-distortion to generative methods with much faster encoding and decoding.
Problem
3DGS compression methods based on implicit neural or customized formats have long encoding/decoding times and are hard to standardize and deploy; traditional point cloud codecs applied directly to 3DGS give poor rate-distortion.
Approach
Generate an explicit, compact 3DGS whose primitives have integer positions, unique coordinates, and quantized low-dimensional latent attributes during training, so the output is directly compatible with standard point cloud encoders (GPCC), with rate control via pruning and bit-depth adaptation.
Key results
- On playroom, HybridGS HR (k_c=3, k_r=2) reaches 29.89 PSNR at 12.15 MB, matching C3DGS PSNR at a smaller size.
- On bicycle, HybridGS LR encodes in 0.66 s and decodes in 0.92 s versus 85.03 s / 80.09 s for HAC.
- Compared with the traditional-compression baseline HGSC, HybridGS is much smaller at similar PSNR (e.g., playroom 28.98 vs 28.83 PSNR at 4.89 MB for CompGS(MM) vs 96.82 MB for HGSC).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase latent color dimension from kc=3 to kc=6 (ablation) | HybridGS · playroom, train, bicycle, room, and dance | PSNR: 29.89 -> 29.89; 21.26 -> 21.49; 24.08 -> 24.10; 29.52 -> 29.75; 39.25 -> 39.31 | Mixed |
| Replace RQ with UQ (ablation) | HybridGS · train, BD 16, High Rate | PSNR: 21.49 -> 21.66 | Helped |
| Replace RQ with UQ (ablation) | HybridGS · dance, BD 16, High Rate | PSNR: 39.31 -> 39.35 | No clear effect |
| Use a trainable low-dimensional latent feature and decoder instead of PCA (ablation) | HybridGS · dance | PSNR: 21.50 -> 39.31 | Helped |
| Replace the nonlinear decoder with a linear decoder (ablation) | HybridGS · dance | PSNR: 39.31 -> 39.27 | Hurt |
| Increase feature bit depth from 12 to 16 (ablation) | HybridGS with RQ · train, High Rate | PSNR: 20.06 -> 21.49 | Helped |
| Increase feature bit depth from 12 to 16 (ablation) | HybridGS with RQ · dance, High Rate | PSNR: 39.23 -> 39.31 | Mixed |
| Remove primitive position uniqueness (ablation) | LQM, BD 10 · room | PSNR: 28.98 -> 30.16 | Mixed |
| Remove primitive position uniqueness (ablation) | UQ, BD 10 · room | PSNR: 30.21 -> 30.23 | No clear effect |
| Use HybridGS HR instead of HAC | 3DGS compression · bicycle | encoding time: 85.03 -> 1.67 | Helped |
| Use HybridGS HR instead of HAC | 3DGS compression · bicycle | decoding time: 80.09 -> 1.77 | Helped |
| Increase latent color dimension from kc=3 to kc=6 (ablation) | HybridGS · all reported scenes | bitstream size | Hurt |
Surprising findings
- Pruning sometimes slightly raises PSNR above the vanilla 3DGS upper bound (e.g., on dance).
- Rate-distortion of HybridGS does not surpass HAC, the best generative method, although it is much faster.
Limitations
- Compression efficiency is lower than end-to-end generative methods trained with RD loss.
- Built on vanilla 3DGS without quality-enhancing modules, so reconstruction quality is capped at vanilla 3DGS.
- Rate control depends on an estimated lossless GPCC ratio, which varies with point density and is less accurate for some settings.
Experimental setup
Datasets: Deep Blending (playroom) Tanks&Temples (train) Mip-NeRF360 (bicycle, room) PKU-DyMVHumans (dance)
Benchmarks: Deep Blending (playroom) Tanks&Temples (train) Mip-NeRF360 (bicycle, room) PKU-DyMVHumans (dance)
Metrics: PSNR file size (MB) encoding/decoding time (s) FPS
Baselines: Scaffold-GS Compact3D C3DGS CompGS(ECCV) LightGaussian Eagles HAC CompGS(MM) GGSC HGSC
Builds on: 3D Gaussian Splatting LightGaussian Robust Quantizer GPCC SparseConv
Closest papers by meaning
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.89)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.89)
- CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression (ICLR 2025 · similarity 0.89)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.89)
- CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting (NeurIPS 2026 · similarity 0.87)
- HybridGS: Decoupling Transients and Statics with 2D and 3D Gaussian Splatting (CVPR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.