Abstract
Feed-forward 3D Gaussian Splatting (3DGS) allows efficient and high-fidelity novel view synthesis (NVS) from an offline recorded image sequence. However, achieving online NVS from streaming and unposed image inputs remains challenging. Although online feed-forward geometric estimation methods have been proposed for streaming depth and point cloud recovery, they cannot be adapted to NVS due to severe rendering artifacts. This is because NVS demands stricter multi-view consistency in Gaussian scales and pose-geometry alignment; even minor deviations would accumulate over time and visibly degrade rendering quality. To this end, we propose FreeStreamGS, a robust online feed-forward framework for efficient and high-quality NVS. We introduce two key mechanisms: a Decoupled Intrinsic Recovery Head that removes cumulative camera intrinsic bias and prevents scene scale jitter during long-term streaming, and a Dynamic Point Refinement Offset strategy that relaxes rigid unprojection to correct coupled pose-depth drift. Extensive experiments show that FreeStreamGS achieves rendering quality competitive with state-of-the-art offline feed-forward 3DGS methods, despite operating without access to future frames.
Method
The pipeline uses causal feature extraction, decoupled camera recovery, Gaussian decoding with DPR-Offsets, and online recursive Gaussian fusion.
Processes frames sequentially with historical context and no future observations.
Decoupled intrinsic recovery reduces scale jitter during long-term streaming.
DPR-Offsets compensate for coupled pose-depth drift before Gaussian fusion.
Quantitative Results
| Dataset | Method | Online | 5-view PSNR | 5-view SSIM | 5-view LPIPS | 64-view PSNR | 64-view SSIM | 64-view LPIPS |
|---|---|---|---|---|---|---|---|---|
| DL3DV-140 | WorldMirror | No | 21.559 | 0.673 | 0.285 | OOM | OOM | OOM |
| DL3DV-140 | OnTheFly-NVS | Yes | 21.030 | 0.644 | 0.325 | 18.478 | 0.610 | 0.455 |
| DL3DV-140 | FreeStreamGS | Yes | 21.884 | 0.688 | 0.273 | 17.190 | 0.558 | 0.485 |
| RE10K | WorldMirror | No | 25.419 | 0.820 | 0.203 | OOM | OOM | OOM |
| RE10K | OnTheFly-NVS | Yes | 20.829 | 0.815 | 0.290 | 20.985 | 0.767 | 0.303 |
| RE10K | FreeStreamGS | Yes | 25.797 | 0.833 | 0.200 | 21.562 | 0.749 | 0.297 |
FreeStreamGS is evaluated under strict online constraints and remains scalable when offline feed-forward baselines run out of memory.
Qualitative Results
Qualitative comparison on DL3DV-140 and RE10K. FreeStreamGS preserves structural detail while operating causally.
Video Stabilization
Once a causal Gaussian representation is available, FreeStreamGS can render stabilized views along smoothed camera trajectories.
BibTeX
@article{chen2026freestreamgs,
title = {FreeStreamGS: Online Feed-forward 3D Gaussian Splatting from Unposed Streaming Inputs},
author = {Chen, Ruiyang and Li, Feiran and Zhou, Chu and Li, Zonglin and Ma, Zhanyu and Guo, Heng},
journal = {arXiv preprint arXiv:2606.03254},
year = {2026}
}