Research map / 3D vision / Gaussian splatting
CodecSplat: Ultra-Compact Latent Coding for Feed-Forward 3D Gaussian Splatting
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 CodecSplat, a learned hierarchical latent codec placed inside a DepthSplat-style feed-forward 3DGS pipeline that entropy-codes the intermediate 2D Gaussian-generation feature, reaching KiB-level scene bitstreams on DL3DV and RE10K.
Problem
Feed-forward 3DGS produces Gaussian primitives without a compact storage/transmission representation, and applying per-scene 3DGS compressors to the generated irregular primitives is inefficient.
Approach
Move the coding bottleneck earlier: entropy-code the dense, view-aligned intermediate Gaussian-generation feature (after depth-guided multi-view refinement) and decode depth and Gaussian attributes from the reconstructed feature, with a two-stage training scheme and a single rate-conditioned codec.
Key results
- On DL3DV, CodecSplat reaches 23.56-26.36 dB PSNR at 20.00-107.77 KiB per scene, versus 1.39 MiB for ReSplat→FCGS at 28.72 dB.
- On RE10K, CodecSplat reaches 24.76-27.05 dB PSNR at 3.37-12.51 KiB per scene, versus 1.63-1.83 MiB for DepthSplat-based compressed baselines.
- Codec-free refiner ablation raises DL3DV PSNR from 25.68 to 26.22 dB; Deflate on float16 features needs 81.08 MiB versus 0.11 MiB for the learned codec at lambda=1024.
Surprising findings
- Naively serializing float16 intermediate features and applying Deflate needs 81.08 MiB per scene, showing a learned codec is necessary even when coding the intermediate representation.
Limitations
- PSNR is capped by the DepthSplat-style backbone and lower-complexity latent-to-Gaussian decoder; lower than ReSplat-based compressed outputs.
- Requires joint training and cannot be plugged into arbitrary pretrained FF3DGS models directly.
- Evaluated only on DL3DV and RE10K with specific view protocols.
Experimental setup
Datasets: DL3DV RealEstate10K
Benchmarks: DL3DV RealEstate10K
Metrics: PSNR SSIM LPIPS bytes per scene timing
Baselines: DepthSplat ReSplat FCGS SOGS gsplat PNG compression
Builds on: DepthSplat hierarchical variational autoencoder FCGS
Closest papers by meaning
- Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis (CVPR 2024 · similarity 0.91)
- Fast Feedforward 3D Gaussian Splatting Compression (ICLR 2025 · similarity 0.91)
- Gaussian Splatting-based Volumetric Video Compression with Sparse 4D Anchors (NeurIPS 2026 · similarity 0.91)
- ContextGS : Compact 3D Gaussian Splatting with Anchor Level Context Model (NeurIPS 2024 · similarity 0.90)
- Optimized Minimal 3D Gaussian Splatting (NeurIPS 2025 · similarity 0.90)
- Efficient Decoupled Feature 3D Gaussian Splatting via Hierarchical Compression (CVPR 2025 · similarity 0.90)
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.