Streaming geometry
LoGeR processes overlapping chunks with hybrid memory, predicting camera poses, dense geometry, and long-context features. A Gaussian decoder maps these features to scene primitives.
Long-context feed-forward 3D reconstruction
L
onGaussian:Tracking Foundation Priors for Long-Context
Feedforward Gaussian Reconstruction
Huazhong University of Science and Technology
* Corresponding author
5.4×
fewer Gaussians than OF3GS
at RE10K / 64 views
3,768
frames in a single-scene
scalability study
Background: LonGaussian rendering along a DL3DV sequence reconstructed from 96 context views.
Explore the workWatch LonGaussian in action
From thousands of input frames to a renderable Gaussian scene.
Download 720p MP443.5 seconds
Long sequences.
Fewer repeated Gaussians.
As a camera moves, the same surfaces appear again and again. Predicting a Gaussian at every input pixel and accumulating these predictions repeatedly represents the same geometry. Streaming the backbone alone does not remove this redundancy.
LonGaussian connects streaming geometry with correspondence-guided consolidation. LoGeR provides long-context features and camera estimates. Within each chunk, multi-frame tracks associate repeated observations, and a learned merger consolidates their Gaussian predictions. A dense complement preserves regions without reliable tracks.
Abstract
Reconstructing high-quality, renderable 3D scenes from long videos is essential for robotic perception and embodied intelligence. Many feed-forward Gaussian reconstruction methods rapidly predict 3D scenes from a small set of images, but extending them to longer sequences presents two bottlenecks: jointly processing all views increases GPU memory requirements, while accumulating pixel-aligned predictions redundantly represents the same surfaces and increases storage and rendering costs.
We present LonGaussian, a streaming feed-forward framework that reconstructs compact, globally renderable Gaussian scenes from long image sequences with known intrinsics and unknown camera poses. Built on a pretrained streaming 3D reconstruction backbone, LonGaussian decodes long-context geometric features into dense Gaussians. Within each chunk, multi-frame point tracks link repeated observations of the same surface and guide a learned merger to consolidate the corresponding Gaussian predictions. Rendered-depth consistency filtering further suppresses Gaussians from other chunks that incorrectly occlude locally reconstructed surfaces. On DL3DV and RealEstate10K benchmarks, LonGaussian improves novel-view rendering quality while using substantially fewer Gaussians than the evaluated feed-forward baselines.
Method
Stream the geometry. Associate observations. Consolidate the representation.
Track-guided consolidation
AllTracker links observations of the same surface within a chunk. The learned merger refines representative Gaussians; a dense complement retains content without reliable track support.
Accumulate and filter
Chunk predictions enter a shared coordinate system. Rendered-depth voting suppresses cross-chunk occluders, while a far-field Gaussian shell models unstable sky geometry.
Visual comparisons
Compare DL3DV and RealEstate10K sequences using one synchronized player. Drag the divider to reveal the selected reference.
DL3DV
Living room · 96 context views
Select a scene and a reference method, then drag to compare.
LonGaussian (Ours)OF3GS
Loading0.0 s
Drag to compare · synchronized video playback
These videos interleave context views (CTX) and held-out novel views (NVS); the badge identifies the current frame. Scenes are reconstructed from context images only. Each method estimates its own camera poses using the paper's evaluation protocol. Playback speed is for visualization and does not measure inference throughput.
Quality and compactness
Benchmark averages using complete accumulated scenes and held-out target views.
| Method | PSNR ↑ | SSIM ↑ | LPIPS ↓ | # Gaussians (M) ↓ |
|---|---|---|---|---|
| ReCoSplat (official) | 13.67 | 0.237 | 0.654 | 13.17 |
| ReCoSplat (adapted) | 15.74 | 0.282 | 0.606 | 13.17 |
| OF3GS | 15.31 | 0.404 | 0.545 | 10.45 |
| LonGaussian (Ours) | 18.88 | 0.575 | 0.420 | 3.54 |
Best entries are emphasized. WorldMirror and AnySplat exceed the 24 GB budget at all reported lengths; see the paper for the full table. ReCoSplat (adapted) is our local resolution adaptation of the released model.
Beyond short sequences
A single-scene case study extending to thousands of frames.
Bounded construction memory
In this scene, construction time and Gaussian count grow approximately linearly across six prefixes. Peak construction memory stays near 20 GB.
Long-range revisits remain challenging: LoGeR pose errors can misalign distant chunks and produce ghosting. The long-sequence image metrics use a local seven-chunk scope, so they do not establish full-scene global rendering fidelity.
Resource measurements cover streaming construction; time excludes model loading, metric rendering, and offline prefix filtering.
Explore the reconstruction
Explore indoor and outdoor scenes reconstructed from complete context sequences. Select a scene, then load its Gaussian rendering.
Gaussian rendering · loads on demand
Drag to orbit · right-drag to pan · scroll to zoom · Reset view to return
Each viewer loads the complete reconstructed scene, with view-dependent Gaussian colors. Explore nearby views within the camera coverage; Reset view returns to the original camera.
Citation
If you find LonGaussian useful, please cite the arXiv paper.
@misc{xie2026longaussiantrackingfoundationpriors,
title={LonGaussian: Tracking Foundation Priors for Long-Context Feedforward Gaussian Reconstruction},
author={Zhoufeng Xie and Guanjun Wu and Shuai Zhang and Xinggang Wang and Wenyu Liu},
year={2026},
eprint={2610.23250},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2610.23250},
}