Towards Geometrically Consistent Novel View Synthesis Using Gaussian Splatting
Achchige, Rusiru Thushara Kuruppu
Achchige, Rusiru Thushara Kuruppu
Supervisor
Department
Computer Vision
Embargo End Date
2025-05-30
Type
Thesis
Date
2025
License
Language
English
Collections
Research Projects
Organizational Units
Journal Issue
Abstract
Novel view synthesis from a single image is a challenging, yet fundamental problem in computer vision and graphics, with applications in mixed reality, gaming, and content creation. While methods such as Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3D-GS) have achieved impressive results, they often require densely captured multi-view images, limiting their practicality in real-world scenarios with sparse or single-image inputs. Recent advances in diffusion models have enabled zero-shot novel view synthesis; however, they struggle with occlusions, lack precise camera control, and often fail to maintain consistency across views. This thesis presents a novel approach for consistency-aware single-view 3D reconstruction by leveraging the complementary strengths of video diffusion models and Gaussian splatting. Video diffusion models, trained on large-scale video datasets, excel at generating plausible motion sequences but lack inherent 3D spatial awareness, making precise viewpoint control difficult. Meanwhile, point cloud-based methods provide a coarse 3D structure but suffer from sparsity, artifacts, and inconsistencies in occluded regions. To address these challenges, we propose an integrated framework that conditions video diffusion priors on explicit 3D information derived from Gaussian splatting, enabling photorealistic rendering with smooth interpolation, view-dependent effects, and structural coherence. By incorporating Gaussian splatting, our method achieves high-fidelity rendering with improved handling of transparency, soft edges, and occlusions compared to traditional point-cloud rendering. Through extensive experiments, we demonstrate that our approach outperforms existing single-view 3D reconstruction methods in terms of visual quality, geometric consistency, and camera pose controllability. This research contributes to the advancement of practical 3D reconstruction of single images, bridging the gap between generative models and explicit 3D representations.
Citation
Rusiru Thushara Kuruppu Achchige, “Towards Geometrically Consistent Novel View Synthesis Using Gaussian Splatting,” Master of Science thesis, Computer Vision, MBZUAI, 2025.
Source
Conference
Keywords
Novel View Synthesis, Video Sysnthesis, 3D Reconstruction