EcoSplat: Efficiency-Controllable Feed-Forward 3D Gaussian Splatting from Multi-View Images

다국어 초록 (Multilingual Abstract)

Feed-forward 3D Gaussian Splatting reconstructs reusable scene representations without per-scene optimization, but dense-view inputs still generate large pixel-aligned primitive sets whose memory and latency costs are hard to predict. EcoSplat makes this efficiency dimension explicit by conditioning reconstruction on a target primitive count. It first learns reliable pixel-aligned Gaussians through Pixel- aligned Gaussian Training (PGT), then performs Importance-aware Gaussian Finetuning (IGF) to rank, suppress primitives under different budgets. EcoSplat preserves high novel-view quality from aggressive 5% budgets to moderate 70% budgets and improves the quality-efficiency tradeoff over prior methods.