WildGaussians logo

WildGaussians

Paid

3D Gaussian Splatting in the Wild

4.5
Type
Saas

About WildGaussians

WildGaussians is a novel 3D Gaussian Splatting (3DGS) approach designed for in-the-wild scenes with appearance variations and dynamic elements. It extends 3DGS by incorporating robust DINO features and an appearance modeling module, enabling real-time rendering while surpassing both 3DGS and NeRF baselines on unconstrained photo collections. Key techniques include per-Gaussian and per-image embeddings fed into an appearance MLP for view-dependent color correction, and a DINO-based uncertainty predictor to remove occluders and floaters. The method achieves state-of-the-art results with inference speeds comparable to or exceeding standard 3DGS, making it suitable for applications like novel view synthesis from unconstrained images.

Key Features

Appearance modeling using per-Gaussian and per-image embeddings with an appearance MLP
DINO-based uncertainty predictor for occluder removal and floater reduction
Real-time rendering speed comparable to or faster than 3DGS (up to 10.6 FPS on NVIDIA 4090 at FullHD)
Smooth interpolation between different scene appearances (e.g., time of day)
Depth prediction via rasterization of Gaussian centers
Simple architectural framework built on 3DGS and Mip-Splatting

Pros & Cons

Pros
  • Achieves state-of-the-art results on in-the-wild data
  • Maintains real-time rendering speed of 3DGS while improving robustness
  • Simple architectural design compared to some NeRF-based approaches
  • Effectively handles both appearance changes and occlusions
Cons
  • Primarily a research project; not available as a commercial product or API
  • Requires DINO features (computation overhead during training)
  • Concurrent works exist (e.g., Wild-GS, SWAG) indicating active competition
  • May require careful tuning for different scene dynamics

Best For

Novel view synthesis from unconstrained photo collections (in-the-wild scenes)Handling images with varying appearance (different times of day, weather)Removing occluders and dynamic distractors from 3D reconstructionsScene appearance interpolation for visualization or editing

Alternatives to WildGaussians

FAQ

What is WildGaussians?
WildGaussians is a research method that extends 3D Gaussian Splatting to handle in-the-wild scenes with varying appearance and dynamic occlusions, using DINO features and an appearance modeling module.
How does WildGaussians handle appearance changes?
It uses per-Gaussian and per-image embeddings passed to an appearance MLP that outputs affine transformation parameters applied to each Gaussian's view-dependent color, enabling interpolation between different scene appearances.
How does it remove occluders?
WildGaussians trains a DINO-based uncertainty predictor that estimates the reliability of each pixel. During rendering, regions with high uncertainty (e.g., occluders) are suppressed, reducing floaters.
What is the rendering speed compared to 3DGS?
WildGaussians matches or exceeds the real-time speed of 3DGS. On an NVIDIA 4090 at FullHD resolution, it achieves 6.5–10.6 FPS depending on the scene, compared to 7.7 FPS for 3DGS.