Research map / 3D vision / Gaussian splatting
GRM: Large Gaussian Reconstruction Model for Efficient 3D Reconstruction and Generation
Yinghao Xu, Zifan Shi, Wang Yifan, Hansheng Chen, Ceyuan Yang, Sida Peng, Yujun Shen, Gordon Wetzstein
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use sigmoid-interpolated scale activation instead of exponential (ablation) | GRM · GSO ablation (256 res) | PSNR: 24.43 -> 27.51 | Improved |
| add transformer upsampler blocks (0 to 1) (ablation) | GRM · GSO ablation (256 res) | PSNR: 27.51 -> 29.11 | Improved |
| increase upsampler blocks from 1 to 3 (ablation) | GRM · GSO ablation (256 res) | PSNR: 29.11 -> 29.38 | Improved |
| add alpha (mask) regularization (ablation) | GRM · GSO ablation (256 res) | PSNR: 29.38 -> 29.48 | Improved |
| replace CNN upsampler with transformer upsampler (ablation) | GRM · GSO ablation (256 res) | PSNR: 27.23 -> 29.48 | Improved |
| predict pixel-aligned depth along rays instead of XYZ coordinates (ablation) | GRM · GSO ablation (256 res) | PSNR: 28.61 -> 29.48 | Improved |
| use LPIPS loss instead of the conventional perceptual loss (ablation) | GRM training · training | training stability | Hurt |
| GRM versus MV-LRM in sparse-view reconstruction with same 4 views | GRM · GSO sparse-view reconstruction (4 views) | PSNR: 25.38 -> 30.05 | Improved |
| GRM versus optimization-based MVDream on text-to-3D retrieval precision | GRM · DreamFusion prompts text-to-3D | R-Prec: 70.1 -> 67.5 | No clear effect |
| GRM speed versus LGM (predicts 16 times fewer Gaussians) | GRM · GSO sparse-view reconstruction (4 views) | Inference time: 0.07 sec -> 0.11 sec | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.
- DreamGaussian — DreamGaussian: Generative Gaussian Splatting for Efficient 3D Content Creation (ICLR 2024) · better on 2 of 2 shared results
- Instant3D — Instant3D: Fast Text-to-3D with Sparse-view Generation and Large Reconstruction Model (ICLR 2024) · better on 1 of 1 shared result
- One-2-3-45++ — One-2-3-45: Any Single Image to 3D Mesh in 45 Seconds without Per-Shape Optimization (NeurIPS 2023) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- GeoLRM: Geometry-Aware Large Reconstruction Model for High-Quality 3D Gaussian Generation (NeurIPS 2024 · similarity 0.90)
- Generative Densification: Learning to Densify Gaussians for High-Fidelity Generalizable 3D Reconstruction (CVPR 2025 · similarity 0.89)
- pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction (CVPR 2024 · similarity 0.88)
- TranSplat: Generalizable 3D Gaussian Splatting from Sparse Multi-View Images with Transformers (AAAI 2025 · similarity 0.88)
- Triplane Meets Gaussian Splatting: Fast and Generalizable Single-View 3D Reconstruction with Transformers (CVPR 2024 · similarity 0.88)
- Generative Sparse-View Gaussian Splatting (CVPR 2025 · similarity 0.88)
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.
