Research map / 3D vision / Gaussian splatting
ZPressor: Bottleneck-Aware Compression for Scalable Feed-Forward 3DGS
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 ZPressor, an architecture-agnostic information-bottleneck-inspired module that compresses dense multi-view features into anchor-view latents via cross-attention, letting feed-forward 3DGS models (DepthSplat, MVSplat, pixelSplat) scale to many input views.
Problem
Feed-forward 3DGS models degrade in quality and run out of memory as the number of input views grows, because pixel-aligned encoders produce linearly more tokens and Gaussians.
Approach
Partition input views into anchor and support sets (anchors chosen by farthest point sampling on camera positions, supports assigned to nearest anchor), then fuse support features into anchor features with stacked cross- and self-attention blocks, so the Gaussian predictor operates on a compact latent state; the compression is framed through the Information Bottleneck principle.
Key results
- On DL3DV with 36 input views, DepthSplat improves from 19.23 to 23.88 PSNR when ZPressor is added.
- On RE10K with 36 views, MVSplat + ZPressor reaches 27.34 PSNR versus 24.19 for MVSplat, while pixelSplat runs OOM without ZPressor.
- ZPressor enables over 100 input views at 480P on an 80GB GPU, per the abstract.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add ZPressor to DepthSplat | DepthSplat · DL3DV, 36 views | PSNR: 19.23 -> 23.88; +4.65 | Helped |
| Add ZPressor to MVSplat | MVSplat · RE10K, 36 views | PSNR: 24.19 -> 27.34; +3.15 | Helped |
| Add ZPressor to pixelSplat | pixelSplat · RE10K, 36 views | PSNR | Helped |
| Use one attention block instead of stacked blocks (ablation) | DepthSplat + ZPressor · DL3DV, 12 views | PSNR: 24.30 -> 24.18 | Hurt |
| Add self-attention (ablation) | DepthSplat + ZPressor · DL3DV, 12 views | PSNR: 23.85 -> 24.30 | Helped |
| Add the IB compression term with beta=10^{-5} rather than beta=0 (ablation) | DepthSplat + ZPressor · DL3DV, 36 views | PSNR: 23.43 -> 23.88 | Helped |
| Use overlap-based anchor selection instead of the default FPS strategy (ablation) | DepthSplat + ZPressor · DL3DV, 36 views | PSNR: 23.88 -> 21.49 | Hurt |
| Use pose-free feature clustering instead of the default FPS strategy (ablation) | DepthSplat + ZPressor · DL3DV, 36 views | PSNR: 23.88 -> 22.81 | Hurt |
| Use K-Means anchor selection instead of the default FPS strategy (ablation) | DepthSplat + ZPressor · DL3DV, 36 views | PSNR: 23.88 -> 22.84 | Hurt |
| Prune views by confidence at ratio 0.75 instead of compressing them with ZPressor | MVSplat · RE10K, 24 views | PSNR: 27.49 -> 21.15 | Hurt |
| Prune views by confidence at ratio 0.5 instead of compressing them with ZPressor | MVSplat · RE10K, 24 views | PSNR: 27.49 -> 26.94 | Hurt |
Surprising findings
- Confidence-based pruning at a 0.75 prune ratio drops MVSplat PSNR from 25.00 to 21.15 on RE10K with 24 views.
Limitations
- Less effective in extremely dense view settings (e.g., 1000 views compressed only to around 50 anchors, still costly).
- Visual quality degrades noticeably with around 500 input images.
- Training uses reduced context views (6 for DepthSplat/MVSplat, 4 for pixelSplat) due to memory limits.
Experimental setup
Datasets: DL3DV-10K RealEstate10K
Benchmarks: DL3DV-10K RealEstate10K ACID
Metrics: PSNR SSIM LPIPS inference time peak memory
Baselines: DepthSplat MVSplat pixelSplat confidence-based pruning
Builds on: DepthSplat MVSplat pixelSplat Information Bottleneck variational information bottleneck
Closest papers by meaning
- CAT-3DGS: A Context-Adaptive Triplane Approach to Rate-Distortion-Optimized 3DGS Compression (ICLR 2025 · similarity 0.91)
- Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis (CVPR 2024 · similarity 0.90)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.90)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.90)
- 3DGS-Enhancer: Enhancing Unbounded 3D Gaussian Splatting with View-consistent 2D Diffusion Priors (NeurIPS 2024 · similarity 0.90)
- Locality-aware Gaussian Compression for Fast and High-quality Rendering (ICLR 2025 · similarity 0.89)
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.