Project Page

Free-Range Gaussians: Non-Grid-Aligned Generative 3D Gaussian Reconstruction

We perform flow matching directly over 3D Gaussian parameters, turning sparse-view reconstruction into a generative process that can synthesize plausible content in unobserved regions.

Ahan Shabanov1,2 Peter Hedman2 Ethan Weber2 Zhengqin Li2 Denis Rozumny2 Gael Le Lan2 Naina Dhingra2 Lei Luo2 Andrea Vedaldi2,3 Christian Richardt2 Andrea Tagliasacchi1 Bo Zhu2 Numair Khan2
1 Simon Fraser University 2 Meta Reality Labs 3 University of Oxford
Sparse Multi-view Input
Sparse input views for object 1.
Sparse input views for object 2.
3D Gaussian Denoising

From a few input images, the model progressively denoises a 3D Gaussian representation.

Abstract

We present Free-Range Gaussians, a multi-view reconstruction method that predicts non-pixel, non-voxel-aligned 3D Gaussians from as few as four images. The model uses flow matching over Gaussian parameters, enabling plausible completions in unobserved regions while avoiding the redundancy of grid-aligned output. To handle larger Gaussian sets efficiently, we introduce a hierarchical patching scheme that groups spatially related Gaussians into joint transformer tokens. We further improve fidelity with timestep-weighted rendering supervision, photometric gradient guidance, and classifier-free guidance.

Method

Sparse-view reconstruction is treated as conditional generation over 3D Gaussian parameters. A diffusion transformer denoises non-grid-aligned Gaussians from noise, while a level-of-detail tree enables coarse-to-fine training and token patchification.

Method overview from the paper.

The core idea is to represent the scene with 3D Gaussians that live directly in space, rather than tying each prediction to a pixel- or voxel-aligned grid structure.

Why Non-Grid-Aligned Generation?

Pixel-aligned methods leave holes in unobserved regions, while voxel-aligned regression tends to blur ambiguous content. The generative formulation produces more complete and sharper reconstructions with far fewer Gaussians.

Non-grid-aligned comparison figure from the paper.

Comparisons

Drag the slider to compare Ours (left, fixed) against baseline methods (right). Switch scenes and select the baseline to inspect failure modes and improvements quickly.

Objaverse
GSO
Input Views
Input views.
Ground Truth
Ground truth views.
Ours
LaRa
Ours image.
Baseline image.
Baseline

Generation vs. Reconstruction

TRELLIS can produce sharp outputs, but often drifts from sparse-view evidence. Free-Range Gaussians better preserves cross-view consistency.

Input Views
Input views.
Ground Truth
Ground truth.
Ours
Our predictions.
TRELLIS
TRELLIS predictions.

BibTeX

@article{free_range_gaussians_2026,
  title   = {Free-Range Gaussians: Non-Grid-Aligned Generative 3D Gaussian Reconstruction},
  author  = {Ahan Shabanov and Peter Hedman and Ethan Weber and Zhengqin Li and Denis Rozumny and Gael Le Lan and Naina Dhingra and Lei Luo and Andrea Vedaldi and Christian Richardt and Andrea Tagliasacchi and Bo Zhu and Numair Khan},
  journal = {arXiv},
  year    = {2026},
  note    = {arXiv ID and link to be added}
}