donydchen/mvsplat

Health v2 · maintenance only

Flags: no_releases

Adoption not part of the score

1296 stars · 84 forks observed · 2026-08-28

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

MVSplat is a PyTorch implementation of an ECCV 2024 Oral model that predicts 3D Gaussians from sparse multi-view images in a single feed-forward pass using a cost volume representation. It achieves state-of-the-art novel view synthesis quality with fast inference (22 fps) and 10x fewer parameters than pixelSplat.

Use cases

• synthesize novel views from a few posed images
• reconstruct 3D scenes as gaussian splats from sparse views
• run feed-forward 3D gaussian splatting inference
• train a multi-view stereo model on RealEstate10K or ACID
• benchmark against pixelSplat for novel view synthesis
• research feed-forward 3D reconstruction with cost volumes

When to choose

• you need fast feed-forward novel view synthesis from sparse multi-view images
• you want a lightweight, efficient alternative to pixelSplat
• you are doing research on cost-volume-based 3D gaussian splatting

When to avoid

• you need per-scene optimization like classic 3D Gaussian Splatting
• you lack a GPU or the training datasets (RealEstate10K/ACID)
• you need high-resolution or 360-degree output without extensions like PanSplat or MVSplat360

Facets

library · maturity active

machine-learning deep-learning image-processing graphics computer-vision graphics machine-learning deep-learning python gaussian-splatting novel-view-synthesis cost-volume feed-forward 3d-reconstruction eccv2024 research-code gpu linux

Member repositories

For agents

markdown · JSON · MCP: product_card(name="donydchen/mvsplat")