DecomVoxel: Harnessing 3D-Native Priors with Guided In-situ Denoising Optimization for Decompositional Scene Reconstruction

2026-10-01

AI Summary

This study addresses the degraded quality of decomposed scene reconstruction under severe occlusion and the spatial drift induced by 3D priors by proposing DecomVoxel. The framework bridges 3D-native generative priors with neural scene reconstruction through guided in-place denoising optimization. It introduces an epsilon distillation loss to stabilize latent variable refinement, combined with adaptive spatial guidance and temporal annealing strategies to suppress artifacts and mitigate drift. Experiments on Replica and ScanNet++ demonstrate that the proposed method significantly outperforms existing state-of-the-art approaches. It achieves high-quality object completion while preserving layout, structural, and textural consistency, ultimately producing high-fidelity meshes.

Abstract

Decompositional scene reconstruction aims to reconstruct high-quality objects and background, yet existing methods still struggle with the level of quality under heavy occlusions. While generative priors offer a potential solution, 2D image-based priors often suffer from multi-view inconsistency due to a lack of 3D awareness. Conversely, 3D-native priors provide stronger structural inductive biases but frequently lead to spatial drift and misalignment within complex scenes. To address these issues, we propose DecomVoxel, formulating object completion as a guided in-situ denoising optimization that bridges 3D-native priors with neural scene reconstruction. Our framework introduces a reformulated epsilon-based distillation loss to ensure stable latent refinement, alongside adaptive spatial guidance that utilizes occupied and vacant anchors with temporal annealing to suppress generative hallucinations and mitigate spatial drift. Experiments on Replica and ScanNet++ show that DecomVoxel significantly outperforms state-of-the-art methods while faithfully preserving the original spatial layout, structural fidelity, and style-consistent texture. Our method pushes the boundary of decompositional reconstruction by delivering high-quality textured meshes with clean topology, geometry, and appearance, providing a robust solution for the decompositional reconstruction of complex real-world scenes. Code is available at https://github.com/DecomVoxel/DecomVoxel.

Problem

• Decompositional Scene Reconstruction
• Heavy Occlusions
• 3D-Native Priors
• Multi-view Inconsistency
• Spatial Drift

Innovation

• Decompositional Scene Reconstruction
• Guided In-situ Denoising Optimization
• 3D-Native Priors
• Epsilon-based Distillation Loss
• Adaptive Spatial Guidance