CuriGS: Curriculum-Guided Gaussian Splatting for Sparse View Synthesis

TL;DR

CuriGS tackles the core challenge of sparse-view 3D reconstruction by introducing curriculum-guided pseudo-views around real camera poses and progressively enlarging viewpoint diversity. The method uses a three-stage pipeline—student view generation, curriculum scheduling, and evaluation/promotion—coupled with depth-correlation and co-regularization to stabilize training. A multi-signal evaluation selects high-quality pseudo-views to augment sparse supervision, yielding improved geometric fidelity and photorealism across LLFF, MipNeRF-360, and DTU, outperforming state-of-the-art baselines under extreme sparsity. The work demonstrates that curriculum-guided augmentation can substantially mitigate overfitting and improve generalization for real-time, explicit Gaussian-splatting representations in sparse-view synthesis.

Abstract

3D Gaussian Splatting (3DGS) has recently emerged as an efficient, high-fidelity representation for real-time scene reconstruction and rendering. However, extending 3DGS to sparse-view settings remains challenging because of supervision scarcity and overfitting caused by limited viewpoint coverage. In this paper, we present CuriGS, a curriculum-guided framework for sparse-view 3D reconstruction using 3DGS. CuriGS addresses the core challenge of sparse-view synthesis by introducing student views: pseudo-views sampled around ground-truth poses (teacher). For each teacher, we generate multiple groups of student views with different perturbation levels. During training, we follow a curriculum schedule that gradually unlocks higher perturbation level, randomly sampling candidate students from the active level to assist training. Each sampled student is regularized via depth-correlation and co-regularization, and evaluated using a multi-signal metric that combines SSIM, LPIPS, and an image-quality measure. For every teacher and perturbation level, we periodically retain the best-performing students and promote those that satisfy a predefined quality threshold to the training set, resulting in a stable augmentation of sparse training views. Experimental results show that CuriGS outperforms state-of-the-art baselines in both rendering fidelity and geometric consistency across various synthetic and real sparse-view scenes. Project page: https://zijian1026.github.io/CuriGS/

Figures (11)

• Figure 1: CuriGS is a curriculum-guided Gaussian Splatting framework that enhances sparse-view 3D reconstruction by progressively introducing pseudo-views with increasing perturbations, yielding stable geometry and photorealistic rendering from extremely limited input views.
• Figure 2: The overall architecture of the CuriGS framework is shown in (A). The pipeline consists of three key stages: (1) student view generation, where pseudo-camera poses are sampled around teacher views with multiple perturbation magnitudes, as detailed in (B); (2) curriculum scheduling, which gradually unlocks perturbation levels during training to progressively expand viewpoint diversity; and (3) student view evaluation and promotion, where each candidate is scored using perceptual (LPIPS), structural (SSIM), and no-reference quality metrics. Only the best student at each perturbation level that passes the evaluation criteria is promoted to the training set, as illustrated in (C). This curriculum-guided process enhances geometric consistency and rendering fidelity under sparse supervision.
• Figure 3: Visualization of student view generation. Examples of pseudo-camera poses with different perturbation magnitudes around teacher view.
• Figure 4: Qualitative comparison on the LLFF dataset. CuriGS achieves sharper details and reduced texture drift compared with other baselines.
• Figure 5: Qualitative comparison on the MipNeRF-360 dataset. CuriGS demonstrates improved perceptual fidelity and geometric stability on large-scale, unbounded scenes.