HIFI-DC: High-fidelity detail-preserving depth completion

Depth completion densifies sparse metric depth (e.g., LiDAR or 3D light-field sensors) into a full map guided by RGB, yet many methods oversmooth thin structures and occlusion boundaries. HiFi-DC is a zero-shot test-time optimization framework that refines an image-conditional latent diffusion prior under three coupled in-loop constraints: (i) weighted least-squares affine metric anchoring to sparse samples, (ii) confidence-weighted sparse supervision to reduce sensitivity to isolated or noisy observations, and (iii) convergence-aware multi-scale normal guidance to encourage structural consistency while avoiding premature over-regularization. A progressive low-to-high-resolution schedule further reduces runtime and allocates late denoising steps to edge refinement. We evaluate HiFi-DC on eight benchmarks (NYU-Depth V2, VOID, KITTI, Hypersim, InteriorNet, MPI-Sintel, ICL-NUIM, and Virtual KITTI 2) using RMSE, MAE, δ 1 , SSIM, and depth-boundary error (DBE). On synthetic datasets with dense and geometrically accurate ground truth, HiFi-DC reduces RMSE by up to 24% and MAE by up to 18% relative to Marigold-DC (e.g., Hypersim RMSE 0 . 698 → 0 . 528 , MAE 0 . 564 → 0 . 463 ), while also improving SSIM and DBE. On real-sensor datasets with sparse or imperfect labels, performance remains competitive in RMSE/MAE, although the benefit of edge-aware refinement is more benchmark-dependent. Overall, the results show that combining metric anchoring and convergence-aware geometric guidance within a frozen diffusion prior can improve detail-preserving depth completion, especially when the supervision is geometrically reliable.