GaMO: Sparse-View 3D Reconstruction with Geometry-Aware Diffusion

Описание: Reconstructing full 3D scenes from just a handful of camera views is one of the hardest problems in computer vision — and GaMO proposes a very elegant solution.

Instead of hallucinating entirely new viewpoints, GaMO (Geometry-aware Multi-view Diffusion Outpainting) takes a different path:
it expands the field of view from existing camera poses using diffusion-based outpainting, while strictly preserving geometric consistency.

Key ideas behind GaMO:
• Works with very sparse views
• Uses Gaussian splatting (3DGS) as the scene representation
• Applies multi-view diffusion outpainting to extend visible regions
• Preserves geometry instead of inventing new camera angles
• Produces wider scene coverage with fewer artifacts

This approach avoids a common failure mode of sparse-view reconstruction, where newly synthesized viewpoints break geometry or drift over time. By growing the scene outward from known camera positions, GaMO keeps everything spatially aligned.

If you’re interested in:
• 3D reconstruction
• Gaussian splats
• Diffusion models beyond text and video
• Geometry-aware generative methods
this project is well worth a closer look.

Project page: https://yichuanh.github.io/GaMO/

GitHub (official): https://github.com/yichuanH/GaMO_offi...

#3DReconstruction #GaussianSplatting #DiffusionModels
#ComputerVision #SparseView #MultiViewGeometry
#NeuralRendering #3DGS #AIResearch
#GenerativeAI #SceneReconstruction #MachineLearning