GeoIn-NeRF: NeRF Inpainting with Geometry-Appearance Prior and Balanced Score Distillation

Recent advances in neural radiance fields (NeRF) inpainting have leveraged pretrained diffusion models to improve object removal and scene completion. However, preserving both visual realism and multi-view geometric consistency remains challenging, as existing diffusion-guided methods often provide insufficient geometric guidance and unstable score-distillation supervision in masked regions. To address these issues, we propose GeoIn-NeRF, a NeRF-specific scene inpainting framework for object removal and scene completion. Specifically, GeoIn-NeRF introduces a Joint Geometric-Appearance Prior (JGAP), which correlates RGB appearance and normal-map geometry in a shared diffusion latent space to provide geometry-aware guidance for NeRF optimization. In addition, it adopts Balanced Score Distillation (BSD) to suppress unstable score components during masked-region optimization. Experiments on representative NeRF inpainting benchmarks show that GeoIn-NeRF achieves favorable visual quality, geometric reconstruction, and subjective preference compared with existing NeRF-based inpainting methods.