Structure-Aware Progressive Refinement for Localized Text-Guided Point Cloud Editing

Localized text-guided point-cloud editing aims to modify a target part according to a natural-language instruction while preserving the geometry of the remaining shape. Existing diffusion-based editors typically rely on text conditions and binary part masks to determine what to edit and where to edit. However, such masks do not explicitly encode the geometric relations that should be maintained within structured parts, often leading to asymmetric components, boundary discontinuities, and multi-scale geometric drift during iterative denoising. We propose STELLAR, a structure-aware progressive refinement framework for localized text-guided point-cloud editing. STELLAR progressively refines the editable region by jointly modeling geometric correspondences among parts, the spatial reliability of point-wise corrections, and multi-scale geometric consistency. The framework comprises three coupled components: part-conditioned symmetry projection, which establishes set-wise geometric correspondences under reflection or rotation priors; dynamic spatial reliability calibration, which uses a source-derived spatial reliability field to attenuate symmetry corrections near fragile attachment boundaries; and hierarchical geometric consistency refinement, which regulates source-relative geometric deviations across scales through residual aggregation and bounded cluster-level corrections. We evaluate STELLAR on localized point-cloud editing across multiple object categories using semantic alignment, structural consistency, shape preservation, and multi-scale geometric stability metrics. Our framework provides a unified mechanism for improving the structural reliability of text-guided local point-cloud edits. Project page: https://stellar-research-site.github.io/.