Adaptive Diffusion Refinement for Enhanced 3D Surface Reconstruction under Geometric Complexity

Abstract Three-dimensional point cloud reconstruction aims to recover continuous, geometrically consistent surfaces from sparse, unstructured samples, playing a pivotal role in 3D perception and modeling systems. Traditional methods often struggle with spatial nonuniformity and abrupt geometric changes, leading to over-smoothing or artifacts. To address these challenges, we introduce a Geometry-Complexity-Aware Diffusion (GCAD) framework that employs diffusion as a region-adaptive feature refiner. This framework utilizes regional complexity scores to drive adaptive timestep allocation and integrates multi-scale features via a confidence-aware fusion module, improving stability and consistency under sparse and noisy inputs. Experiments on ShapeNet, SyntheticRoom, and ScanNet demonstrate consistent improvements over strong baselines, with GCAD achieving a 15\% reduction in Chamfer Distance and a 10\% increase in F-score on ShapeNet. The adaptive scheduling mechanism also ensures a better balance between quality and computational efficiency. Our findings substantiate the effectiveness of geometry-complexity-aware adaptive refinement in enhancing reconstruction accuracy and geometric consistency.Our source code are available at https://github.com/kamiya1001hana/GCAD.git