CompactSplat allocates 3D Gaussians by scene content
Original title: CompactSplat: Spatially Adaptive Gaussian Distribution for Feedforward Scene Reconstruction
The feedforward method uses fewer primitives and can change sparsity at inference without retraining.
CompactSplat is a feedforward 3D Gaussian Splatting framework that places primitives by scene content instead of assigning one or more Gaussians to every pixel. Texture-guided partitioning and hierarchical sampling pack dense Gaussians into complex regions and larger primitives on flat surfaces. On RealEstate10K, DL3DV, and ScanNet, the authors report better metrics and high-resolution consistency with fewer primitives.
- Uniform per-pixel Gaussians waste primitives on homogeneous areas.
- Inference-time sparsity control trades fidelity against memory without retraining.
- Experiments cover RealEstate10K, DL3DV, and ScanNet.
From the original article
1 min read at huggingface.co
Reconstructing 3D scenes from sparse views without per-scene optimization remains highly challenging, especially for recovering accurate geometry and fine textures. While recent feedforward approaches leverage generalizable 3D Gaussian Splatting (3DGS) for scene generation, they typically assign one or multiple Gaussians to each pixel.
By Bing He, Jingnan Gao, Yunuo Chen · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.