3D Gaussian Splatting (3DGS) has recently enabled real-time rendering of unbounded 3D scenes for novel view synthesis. However, this technique requires dense training views to accurately reconstruct 3D geometry. A limited number of input views will significantly degrade reconstruction quality, resulting in artifacts such as “floaters” and “background collapse” at unseen viewpoints. In this work, we introduce SparseGS, an efficient training pipeline designed to address the limitations of 3DGS in scenarios with sparse training views. SparseGS incorporates depth priors, novel depth rendering techniques, and a pruning heuristic to mitigate floater artifacts, alongside an Unseen Viewpoint Regularization module to alleviate background collapses. Our extensive evaluations on the Mip-NeRF360, LLFF, and DTU datasets demonstrate that SparseGS achieves highquality reconstruction in both unbounded and forwardfacing scenarios, with as few as 12 and 3 input images, respectively, while maintaining fast training and real-time rendering capabilities.

The challenge of learning 3D representations from 2D images has been a longstanding area of interest, but achieving a balance between efficiency and fidelity remains a persistent challenge. While Neural Radiance Fields (NeRFs) excel in high-quality rendering and effectively represent anisotropic effects in view interpolation, they suffer from long training times, and blurriness if only sparse views are provided as input. The recent development of 3D Gaussian Splatting (3DGS) has substantially reduced the training cost by introducing a more compact, explicit 3D representation coupled with a real-time rendering pipeline. However, 3DGS still suffers from artifacts caused by the inherent ambiguity in projection from 3D to 2D posed by sparse input views.

Artifacts in 3DGS, such as “floaters” (high-density floating regions due to misplaced Gaussians) and “background collapse” (caused by Gaussians being misplaced at incorrect depths, resulting in background Gaussians appearing in the foreground) tend to be more pronounced than those in NeRFs. These issues are further exacerbated when the training set lacks substantial scene coverage, such as in multi-view unbounded scenes [2] (referred as 360-degree scenes in the rest of this paper). Extensions of 3DGS, such as [14, 49], have attempted to incorporate depth priors as geometry supervisions or regularizers, but they still fail to arXiv:2312.00206v4 [cs.CV] 23 Jun 2026 resolve the problem of floaters, particularly in unbounded scenes. We observe that naively applying the same alphablending equation for rendering depth can cause gradients to propagate to the wrong Gaussians, adversely affecting quality. Additionally, only providing extra guidance from the training views does not mitigate the problem of overfitting, thus failing to address background collapse issues in sparse-input settings.

We conduct our experiments on three datasets, categorized into two settings: 1) the Mip-NeRF360 [2] dataset, which features seven challenging 360° scenes; 2) the LLFF [19] and DTU [10] datasets, which contain forward-facing scenes. For the Mip-NeRF360 dataset, we use every eighth image as testing view and evenly sample 12 or 24 views from the remaining views as the training set. For the LLFF and DTU datasets, we follow the training and evaluation protocols established in RegNeRF [24]. Following conventions, all input images are downscaled to 1/4 of the original height and width.

In this paper, we propose a method using 3D Gaussian Splatting (3DGS) representation to tackle sparse-view 3D reconstruction task. We observe that the alpha blending depth rendering in original 3DGS often results in misplaced Gaussians (known as “floaters”). Therefore, we propose to constrain depth convergence with softmax-scaling and mode-selection bias that significantly reduce such floaters. In regions with little coverage by input views, we leverage Score Distillation Sampling (SDS) and Depth Warping to reduce collapse in geometry and noise in texture while preserving fine details. Lastly, we propose a novel floater pruning process to identify and remove low-opacity floaters. In evaluation, we show that our method outperforms the stateof- the-art methods by outputting a much cleaner and more coherent scene under even more challenging 12-view setting on Mip-NeRF360 dataset.