Research map / 3D vision / Gaussian splatting
GSDF: 3DGS Meets SDF for Improved Neural Rendering and Reconstruction
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 GSDF, a dual-branch 3DGS + neural SDF framework with mutual guidance (depth-guided SDF ray sampling, SDF-guided Gaussian density control, and depth/normal consistency) to improve both rendering and surface reconstruction.
Problem
Neural rendering and surface reconstruction conflict: constraints that help geometry (flattened Gaussians, binary opacity, strict density fields) degrade rendering quality, and SDF-based methods converge slowly and struggle with fine detail.
Approach
Jointly train a Scaffold-GS-based Gaussian branch and a NeuS/hash-grid SDF branch where each guides the other during training: rasterized GS depth sets SDF ray sampling ranges, SDF values steer Gaussian growing and pruning, and depth and normal maps from both branches are aligned, without constraining Gaussian shapes.
Key results
- On DTU, GSDF achieves the best Chamfer distance (0.802) versus NeuS (0.823) and 2D-GS (0.928), with PSNR 33.65.
- On Mip-NeRF360, GSDF reaches PSNR 29.38 and LPIPS 0.185, improving over Scaffold-GS (29.34 / 0.200) in the reported table.
- With random Gaussian initialization, GSDF (rand) outperforms Scaffold-GS (rand) on Mip-NeRF360 (28.05 vs 27.84 PSNR as averaged across table entries).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add depth-guided ray sampling to the SDF branch (ablation) | GSDF · Mip-NeRF360, Tanks&Temples, Deep Blending | PSNR: 29.26 -> 29.38; 27.30 -> 27.37; 30.17 -> 30.38 | Helped |
| Add geometry-aware Gaussian density control (ablation) | GSDF · Mip-NeRF360, Tanks&Temples, Deep Blending | PSNR: 29.29 -> 29.38; 27.29 -> 27.37; 30.25 -> 30.38 | Helped |
| Add mutual geometric supervision (ablation) | GSDF · Mip-NeRF360, Tanks&Temples, Deep Blending | PSNR: 29.27 -> 29.38; 27.18 -> 27.37; 30.21 -> 30.38 | Helped |
| Use GSDF rather than Scaffold-GS with random Gaussian initialization | Gaussian rendering · Mip-NeRF360, Tanks&Temples, Deep Blending | PSNR: 27.84 -> 28.05; 26.37 -> 26.55; 29.12 -> 29.57 | Helped |
Surprising findings
- Ablating mutual geometric supervision degrades surface quality and rendering details, which the authors attribute to neural rendering being more tolerant of deviations than surface reconstruction.
Limitations
- Struggles with reflections and intense lighting variations, such as indoor scenes.
- Higher memory usage and longer training since two representations are optimized jointly.
- The SDF branch (MLP-based) lags behind the GS branch in efficiency.
Experimental setup
Datasets: DTU Mip-NeRF360 Tanks&Temples Deep Blending
Benchmarks: DTU Mip-NeRF360 Tanks&Temples Deep Blending
Metrics: PSNR SSIM LPIPS Chamfer distance
Baselines: NeuS Instant-NSR 3D-GS Scaffold-GS 2D-GS SuGaR
Builds on: 3D Gaussian Splatting Scaffold-GS NeuS Instant-NSR PermutoSDF
Closest papers by meaning
- Neural Signed Distance Function Inference through Splatting 3D Gaussians Pulled on Zero-Level Set (NeurIPS 2024 · similarity 0.93)
- Gaussian Herding across Pens: An Optimal Transport Perspective on Global Gaussian Reduction for 3DGS (NeurIPS 2025 · similarity 0.90)
- GaussianUDF: Inferring Unsigned Distance Functions through 3D Gaussian Splatting (CVPR 2025 · similarity 0.90)
- Evolving High-Quality Rendering and Reconstruction in a Unified Framework with Contribution-Adaptive Regularization (CVPR 2025 · similarity 0.90)
- MS-GS: Multi-Appearance Sparse-View 3D Gaussian Splatting in the Wild (NeurIPS 2025 · similarity 0.89)
- GVKF: Gaussian Voxel Kernel Functions for Highly Efficient Surface Reconstruction in Open Scenes (NeurIPS 2024 · 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.