Title: FreeStreamGS: Real-Time Feed-Forward 3D Gaussian Splatting Using Unposed Streaming Data

Abstract: Feed-forward 3D Gaussian Splatting (3DGS) has established itself as a method for achieving high-fidelity novel view synthesis (NVS) efficiently, provided the input is an offline-recorded sequence of images. However, performing online NVS using streaming inputs that lack camera pose information remains a significant hurdle. While researchers have developed online feed-forward techniques for recovering streaming depth and point clouds, these approaches are ill-suited for NVS because they produce unacceptable rendering artifacts. This limitation stems from the fact that NVS requires strict multi-view consistency regarding Gaussian scales and precise alignment between pose and geometry; even slight errors tend to compound over time, leading to noticeable declines in visual quality. To address these challenges, we introduce FreeStreamGS, a robust framework designed for efficient, high-quality online feed-forward NVS. Our approach incorporates two primary innovations: a Decoupled Intrinsic Recovery Head, which eliminates cumulative bias in camera intrinsics and stabilizes scene scale during extended streaming sessions, and a Dynamic Point Refinement Offset strategy, which mitigates coupled pose-depth drift by moving away from rigid unprojection. Our extensive experimental results demonstrate that FreeStreamGS delivers rendering quality comparable to leading offline feed-forward 3DGS methods, all while operating without the benefit of future frames.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC