NVlabs/InstantSplat

Health v2 · maintenance only

Flags: no_releases no_license

Adoption not part of the score

1694 stars · 156 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-30, confidence not recorded

InstantSplat is a research framework for photorealistic 3D scene reconstruction from extremely sparse image views using Gaussian Splatting, eliminating the need for Structure-from-Motion. It combines feed-forward geometric priors (MASt3R/DUSt3R) with differentiable neural rendering to jointly estimate camera poses and 3D Gaussians in seconds, over 30x faster than COLMAP-based pipelines.

Use cases

• reconstruct 3D scenes from a few photos
• novel view synthesis from sparse images
• estimate camera poses without COLMAP
• render free-viewpoint videos of a scene
• fast 3D gaussian splatting reconstruction
• photorealistic scene capture for robotics or graphics research

When to choose

• you have only 2-3 images of a scene and need a 3D reconstruction quickly
• SfM tools like COLMAP fail or are too slow for your pipeline
• you want a research baseline for sparse-view Gaussian Splatting with 3D-GS or 2D-GS support

When to avoid

• you need production-grade, licensed software (license is non-standard)
• you lack a CUDA-capable GPU or cannot build custom CUDA kernels
• you need long-sequence multi-window alignment, which is still on the TODO list

Facets

library · maturity active

graphics machine-learning image-processing simulation computer-vision graphics machine-learning artificial-intelligence python gaussian-splatting 3d-reconstruction novel-view-synthesis sparse-view sfm-free neural-rendering pose-estimation research-code linux gpu docker

Member repositories

For agents

markdown · JSON · MCP: product_card(name="NVlabs/InstantSplat")