| license: mit | |
| language: | |
| - en | |
| tags: | |
| - computer-vision | |
| - 3d-reconstruction | |
| - novel-view-synthesis | |
| - gaussian-splatting | |
| - diffusion-models | |
| - sparse-view-reconstruction | |
| library_name: pytorch | |
| pipeline_tag: unconditional-image-generation | |
| # GeoQuery Checkpoint | |
| This repository provides the pretrained checkpoint for **GeoQuery**, a geometry-guided diffusion framework for sparse-view 3D scene reconstruction and novel view refinement. | |
| GeoQuery is designed to improve the rendering quality of sparse-view 3D Gaussian Splatting by using diffusion priors and explicit geometric guidance. Given artifact-prone novel-view renderings and reference views, GeoQuery performs geometry-aware restoration to reduce floating artifacts, structural distortions, texture errors, and cross-view inconsistencies. | |
| > Code repository: [GeoQuery](https://github.com/Xiaoc7/GeoQuery) | |
| ## Model Description | |
| GeoQuery introduces a geometry-guided query mechanism for diffusion-based novel view refinement. Instead of relying purely on corrupted rendered-view features for cross-view information retrieval, GeoQuery leverages explicit geometric correspondences to guide feature aggregation from reference views. | |
| The released checkpoint is intended for research usage and reproducing the results of the GeoQuery project. | |
| ## Checkpoint | |
| The checkpoint file included in this repository: | |
| ```text | |
| geoquery_win3_down1.pkl | |