EcoSplat:

Efficiency-controllable Feed-forward 3D Gaussian Splatting from Multi-view Images

Jongmin Park*      Minh-Quan Viet Bui*      Juan Luis Gonzalez Bello      Jaeho Moon      Jihyong Oh†      Munchurl Kim†
*Co-first authors (equal contribution)
†Co-corresponding authors
KAIST, South Korea        Chung-Ang University, South Korea
arXiv Code

NVS under Controlled Number of Primitives

We compare our EcoSplat with other representative feed-forward 3DGS methods under varying target 3D Gaussian counts, which are set to 5%, 10%, 40%, and 70% of the total pixel-aligned primitives (NxHxW).

RE10K Dataset Comparisons

ACID Dataset Comparisons

Abstract

TL;DR: we propose EcoSplat, the first efficiency-controllable feed-forward 3DGS framework that adaptively predicts the 3D representation for any given target primitive count at inference time. Extensive experiments across multiple dense-view settings show that EcoSplat is robust and outperforms state-of-the-art methods under strict primitive-count constraints, making it well-suited for flexible downstream rendering tasks.

teaser

Framework Architecture

architecture

EcoSplat Overview. EcoSplat is trained in two stages: Pixel-aligned Gaussian Training (PGT) (Sec. 3.2) and Importance-aware Gaussian Finetuning (IGF) (Sec. 3.3). During IGF, the combination of the importance-aware opacity loss and the Progressive Learning on Gaussian Compaction (PLGC) encourages EcoSplat to suppress the opacities of less important Gaussians. At inference, it adaptively satisfies an arbitrary user-specified primitive count and produces the optimal Gaussians in a feed-forward manner (Sec. 3.4).