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

2026-04-06

AI Summary

Existing generative 3D Gaussian methods are constrained by pixel- or voxel-grid alignment, often leading to redundancy, voids, and ambiguity in unobserved regions. This work proposes a non-grid-aligned generative framework for 3D Gaussian reconstruction that directly predicts Gaussian parameters in continuous space via flow matching. To mitigate the computational burden of long sequences, a hierarchical Gaussian grouping Transformer is introduced. The approach further enhances reconstruction fidelity through timestep-weighted rendering loss, photometric gradient guidance, and classifier-free guidance. Evaluated on Objaverse and Google Scanned Objects, the method achieves high-quality reconstructions with fewer Gaussians, demonstrating particularly strong performance under sparse input or missing multi-view conditions.

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. This is done through flow matching over Gaussian parameters. Our generative formulation of reconstruction allows the model to be supervised with non-grid-aligned 3D data, and enables it to synthesize plausible content in unobserved regions. Thus, it improves on prior methods that produce highly redundant grid-aligned Gaussians, and suffer from holes or blurry conditional means in unobserved regions. To handle the number of Gaussians needed for high-quality results, we introduce a hierarchical patching scheme to group spatially related Gaussians into joint transformer tokens, halving the sequence length while preserving structure. We further propose a timestep-weighted rendering loss during training, and photometric gradient guidance and classifier-free guidance at inference to improve fidelity. Experiments on Objaverse and Google Scanned Objects show consistent improvements over pixel and voxel-aligned methods while using significantly fewer Gaussians, with large gains when input views leave parts of the object unobserved.

Problem

• 3D Gaussian reconstruction
• non-grid-aligned
• multi-view reconstruction
• unobserved regions
• redundant Gaussians

Innovation

• non-grid-aligned
• generative 3D Gaussian
• flow matching
• hierarchical patching
• classifier-free guidance