KoalaGS: Complexity-Aware Continuous Level of Detail for Gaussian Splatting

TL;DR

KoalaGS introduces a complexity-aware continuous level of detail method for 3D Gaussian Splatting that dynamically reallocates Gaussian primitives during training based on precomputed image complexity. This ensures smoother degradation across rendering budgets and preserves scene structure at low capacities. The approach improves quality across all budget levels and enables faster progressive streaming.

Why It Matters

It enables efficient deployment of 3D scenes on devices with limited compute by maintaining visual fidelity across varying resource constraints. This is especially valuable for real-time applications like AR/VR and mobile rendering, where smooth transitions and fast streaming are critical.

Key Findings

• Image complexity computed from input views before optimization predicts reconstruction error hotspots in both 2D and 3D space.
• Simple regions are over-represented with Gaussians, leading to disproportionate degradation under budget constraints.
• Using a complexity-driven importance score during training enables smooth, structure-preserving degradation across continuous levels of detail.