HierGF: Hierarchical Gaussian Fields via Geometry-perception Message Passing for Sparse-view 3D Reconstruction

The two-stage hierarchical backbone with selective inheritance is conceptually clean and well justified: by not transferring the full coarse state, the fine field retains freedom to refine shape and appearance under stronger pseudo-supervision. The geometry-consistent densification module is a concrete improvement over standard 3DGS adaptive density control, using voxel-based sparsity detection and normal-weighted interpolation to place new Gaussians along surfaces rather than at arbitrary locations. The ablation in Table XIII clearly shows that selective transfer outperforms both full reinitialization and transferring all attributes.

The paper makes strong causal claims about geometry-perception message passing, yet the largest performance boosts appear inseparable from the addition of a 2D diffusion prior and monocular depth supervision (ZoeDepth), both of which are external signals rather than self-generated geometry. While the confidence network attempts to filter diffusion artifacts, its training relies on hand-designed regularizers—$\mathcal{L}_{\text{conf}} = \lambda_{\text{view}}\sum_{j\in N^{\prime}}(\mathbf{w}_{j}-1)^{2} + \lambda_{\text{tv}}\sum_{j\in N^{\prime}}\operatorname{TV}(\mathbf{M}_{j}) + \lambda_{\text{pix}}\sum_{j\in N^{\prime}}(\overline{\mathbf{M}}_{j}-1)^{2}$—that encourage weights near one and smooth maps, which may simply dilute unreliable signals rather than truly learn reliability.\nThere is also no quantitative analysis of geometry quality independent of rendering metrics; PSNR and SSIM are view-synthesis metrics, so reported gains could reflect better texture hallucination rather than more accurate 3D structure. The densification strategy depends on hyperparameters such as voxel size $s_{v} = 0.02 D_{\text{rob}}$ and neighborhood radius $r = 0.04 D_{\text{rob}}$ whose sensitivity is only briefly ablated on two datasets, leaving open how they generalize to scenes with drastically different scales or thin structures.

Quantitative comparisons span five benchmarks and multiple view counts, which is commendable, but the inclusion of older or mismatched baselines (e.g., DVGO, DietNeRF) alongside recent Gaussian methods somewhat dilutes the focus. The paper is transparent about using reported versus re-run results and about differing priors across methods (Table VII), yet direct comparisons with the most related diffusion-assisted baselines—3DGS-Enhancer and IPSM-Gaussian—are limited to qualitative or conceptual discussion rather than head-to-head numbers on shared splits. On LLFF and DTU, HierGF leads by small margins over NexusGS and CoR-GS, but the gains are within roughly 0.2–0.5 dB PSNR, raising questions about statistical significance given that only a single run per scene is mentioned. The qualitative figures look favorable, yet without error maps or user studies they remain subjective.

The authors provide unusually detailed implementation settings in Table V, including optimization budgets (4,000 coarse + 6,000 fine iterations), LoRA adaptation steps (500), pseudo-view refresh frequency ($K=500$), SfM filtering thresholds, and exact input resolution ($779 \times 520$). These specifics should enable close replication if code is released. Unfortunately, no code or pre-trained weights are mentioned, and the diffusion adaptation requires a frozen ControlNet/Stable Diffusion v1.5 backbone plus LoRA parameter updates, which introduces dependencies on diffusers/checkpoint versions that are not specified. The runtime of 632 s is given for a single scene on an A100 GPU but without standard deviations or scene averages, making it hard to assess variance. Reproducing the geometry-consistent densification also hinges on the ZoeDepth prior and the SfM initialization quality from the sparse input views alone, which is a known brittle ingredient in sparse-view pipelines.