Diffusion-Based 3D Reconstruction for Geometry Consistency from a Single Image

Single-image 3D reconstruction remains challenging due to unreliable geometry and inconsistent camera pose estimation. We introduce a diffusion-based pipeline that converts a single input image into a geometrically consistent pseudo-multiview set suitable for stable 3D Gaussian Splatting. Although SV3D synthesizes diverse viewpoints its outputs often contain geometric distortions and unstable view trajectories. To address these issues we apply a three-stage preprocessing and geometric initialization pipeline: (1) appearance-consistency–based filtering along the synthesized view trajectory to select structurally reliable frames, (2) a super-resolution step that restores high-frequency texture details, and (3) VGGT-guided geometric initialization that enables robust COLMAP pose estimation. This pipeline transforms raw SV3D sequences into a coherent pseudo-multiview dataset with stable viewpoints, enhanced textures, and consistent geometric cues. As a result, our method achieves more reliable camera registration and improved reconstruction quality in terms of geometric consistency and visual stability compared to existing single-view approaches relying on strong prior-driven geometry.