IE-SRGS: An Internal-External Knowledge Fusion Framework for High-Fidelity 3D Gaussian Splatting Super-Resolution

TL;DR

IE-SRGS addresses the core challenge of high-fidelity 3D Gaussian Splatting super-resolution from low-resolution inputs by fusing external 2D SR priors with internal 3DGS representations. It introduces a mask-guided fusion strategy that leverages external texture and depth priors alongside an internal multi-scale 3DGS backbone with MV-Regulation, producing cross-view-consistent and detail-rich HR reconstructions. Extensive experiments on synthetic and real datasets show state-of-the-art quantitative gains and qualitative improvements, closely approaching HR upper bounds while maintaining efficiency. The work demonstrates the viability and benefits of joint internal–external knowledge fusion for 3D SR and establishes a foundation for future unified 3D low-level tasks.

Abstract

Reconstructing high-resolution (HR) 3D Gaussian Splatting (3DGS) models from low-resolution (LR) inputs remains challenging due to the lack of fine-grained textures and geometry. Existing methods typically rely on pre-trained 2D super-resolution (2DSR) models to enhance textures, but suffer from 3D Gaussian ambiguity arising from cross-view inconsistencies and domain gaps inherent in 2DSR models. We propose IE-SRGS, a novel 3DGS SR paradigm that addresses this issue by jointly leveraging the complementary strengths of external 2DSR priors and internal 3DGS features. Specifically, we use 2DSR and depth estimation models to generate HR images and depth maps as external knowledge, and employ multi-scale 3DGS models to produce cross-view consistent, domain-adaptive counterparts as internal knowledge. A mask-guided fusion strategy is introduced to integrate these two sources and synergistically exploit their complementary strengths, effectively guiding the 3D Gaussian optimization toward high-fidelity reconstruction. Extensive experiments on both synthetic and real-world benchmarks show that IE-SRGS consistently outperforms state-of-the-art methods in both quantitative accuracy and visual fidelity.

Figures (8)

• Figure 1: Comparison between existing methods and IE-SRGS. Left: External 2D Super-Resolution (2DSR) models provide detail but lack consistency; internal 3D models offer consistency but lack details. IE-SRGS achieves consistent and detailed outputs. Right: Existing methods rely solely on external priors. IE-SRGS jointly integrates external and internal knowledge, enabling high-quality high-resolution (HR) 3D Gaussian Splatting (3DGS) from low-resolution (LR) multi-view inputs.
• Figure 2: Overall framework of the proposed IE-SRGS. Given LR multi-view images, IE-SRGS extracts external guidance from pre-trained 2DSR models and internal guidance from a multi-scale 3D model. A mask-guided internal-external guidance integration provides structured, adaptive supervision for final HR 3DGS optimization.
• Figure 3: Qualitative comparisons of 4$\times$ 3D super-resolution on the NeRF Synthetic dataset. IE-SRGS achieves sharper textures and higher fidelity compared to SOTA methods 3DGS, Mip-Splatting, SwinIR-3DGS, and SRGS.
• Figure 4: Qualitative comparisons of 4$\times$ 3D super-resolution on real-world datasets. IE-SRGS reconstructs finer textures and preserves more structural details compared to SOTA methods 3DGS, Mip-Splatting, SwinIR-3DGS, and SRGS.
• Figure 5: Qualitative results of the component effectiveness analysis. MV-Regulation is first added to Mip-Splatting to form a consistency baseline, while external, internal, and joint guidance components are progressively added without MV-Regulation.