- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
PBE augments the prediction head of a frozen feed-forward reconstruction backbone with spatial Beta and three-lobe Spherical Beta parameters. It constructs heuristic geometry-based single-view bases from depth and normals, combines multi-view evidence with attribute-specific gates, and predicts learned residual corrections. The authors train the augmented heads on DL3DV and evaluate Depth Anything 3, VGGT-Ω, AnySplat and YoNoSplat on DL3DV, Mip-NeRF360 and Tanks and Temples. Table 1 reports PSNR gains across the tested backbone–dataset combinations.
Critical strengths
The approach operates at the primitive-parameter prediction stage, retaining frozen feature backbones and feed-forward inference. Demonstrating the extension across four architectures supports its potential practical reuse rather than dependence on one particular reconstruction backbone. Table 1 reports improved PSNR for all four architectures on all three benchmarks, often with the 50K-step extension exceeding a 100K-step original-head baseline. This is useful evidence of improved reconstruction performance under the tested settings.
Critical weaknesses
-
The spatial Beta base relies on a hand-weighted combination of normalized depth gradients and a curvature proxy (Eq. 2), while SB directions and initial sharpness use fixed angle and scale choices. These may be useful inductive biases, but are not clearly derived from physical image formation. Table 2 shows just 0.23 dB over a random-prior variant, while the full comparison also changes kernels, appearance basis, prediction head and multi-view evidence.
-
Table 3 shows very similar scores for Single and Full at a fixed checkpoint; Single still retains multi-view evidence in the residual branch, so this is not a clean independently trained fusion ablation. In Table 1, YoNoSplat gains PSNR but substantially worsens LPIPS on all three benchmarks, an important caveat for the claimed improvements in specular appearance.
Questions for authors
Q1: Would a Gaussian/SH head with a similarly sized residual prediction module and the same geometry/multi-view inputs reproduce much of the gain? Please separate benefits of richer Beta/SB expressiveness from extra head capacity.
Q2: Can independently retrained ablations isolate spatial Beta, SB appearance, geometric initial bases and multi-view base fusion? A frozen-checkpoint intervention does not determine how a different model would perform after retraining.
Q3: What causes the repeatable YoNoSplat LPIPS degradation despite improved PSNR? Also, the abstract states a DA3 Mip-NeRF360 baseline of 24.25 dB, but Table 1 lists 24.55 dB; are these different experimental settings or a reporting error?
2 comments
Thank you for your suggestions on our work. For W2, in Fig. 7 of Appendix C, we further present the rendering comparison of YoNoSplat on an example view (000001), and we zoom in on some regions that are difficult to render. The results demonstrate the improvements of our PBE method for these three algorithms. In terms of PSNR, PBE brings improvements of dB for YoNoSplat, $1.776 \Omega$, and dB for AnySplat, respectively.
For W1: Thank you for this observation. As for these manual settings, the explicit basis provides input-dependent parameter references, enabling the network to learn corrections to the reference values. Setting an initial configuration facilitates parameter optimization. For the SB direction and initial sharpness, we use fixed angle and ratio choices: three directions at different orientations provide dispersed initial angular responses, and 2, 4, and 8 provide initial responses from wide to narrow. These are initialization choices, not constants derived from physical laws. Through the initial settings and the constraints of the loss, we expect the SB directions to cover the surface of the primitives rather than collapse entirely into a single direction. The random prior results indicate that the manual settings have little effect on the overall performance, although they may cause larger errors in training with only a few epochs. In Fig. 5 of Appendix B, we provide a loss comparison between PBE and the random prior; the PBE loss is lower than that of the random prior in every epoch, and the loss gap is very clear in the first 5 epochs, with a maximum gap of about 0.1. We provide a more detailed ablation in Table 4, which demonstrates that our PBE method is effective and that the multi-view evidence, kernels, and prediction head are all useful.
For W2 and Q3:Thank you for this suggestion. We provide a more detailed ablation in Table 4, which demonstrates the effectiveness of the multi-view evidence. We ablated the multi-view features from the features, and all metrics on the three datasets were lower than those of PBE. Regarding the LPIPS degradation issue, in the current Table 2 we further added comparative experiments on Re10K (68): YONOSPLAT's LPIPS did not degrade, whereas AnySplat's LPIPS degraded. The reason is that our loss consists of an L1 loss and a gradient loss; during training, the model may focus too much on pixel-wise error and lack supervision of some local structures. Second, we did not tune different parameters for different methods during training; different architectures may respond differently to the same training configuration, which can also lead to LPIPS degradation. The specular reflections we mention actually refer to cases in some scenes where sudden highlights may appear. Low-order SH coefficients are limited in representing local, sharp view-dependent angular responses, and our PBE method can compensate for this deficiency. For example, in the top-left image of Fig. 1, the PBE method recovers most of the black regions of the water-surface reflection, while the original DA3 is clearly worse than PBE. The abstract states that the DA3 Mip-NeRF360 baseline is 24.25 dB, whereas Table 1 lists 24.55 dB. This is because in the abstract we report the result after 5 epochs of training, which is 24.25 dB, while Table 1 is after 10 epochs, which is 24.55 dB.
For Q1:Thank you for this suggestion. In Table 4, on DA3 we used exactly the same residual and geometric construction as PBE, as well as a residual-only module; the module parameter count, geometric construction method, and multi-view components are exactly the same as PBE. We then trained them separately on DL3DV for 5 epochs. As shown in the table, the metrics on all three test sets drop substantially: PSNR drops by about 11 dB, and SSIM and LPIPS drop by about 0.3 and 0.5, respectively, compared with the baseline. This indicates that the additional residual prediction does not provide a gain; the main gain comes from Beta/SB. Moreover, in this part of the experiment, the Gaussian attributes with the added geometric construction achieved better results than residual optimization alone, further demonstrating the effectiveness of the geometric construction.
For Q2:Thank you for this suggestion. In Table 4, we further discuss the expressiveness of Beta/SB. Combining either the beta parameter alone or SB alone with Gaussian primitives both achieves better results than the original Gaussian. w/o spatial Beta improves clearly over the DA3 baseline: PSNR increases by about 1.51–1.83 dB, SSIM increases by about 0.0389–0.0608, and LPIPS decreases by about 0.0345–0.0774. w/o SB improves less over the DA3 baseline: PSNR increases by 0.26/0.24 dB on DL3DV/Mip-NeRF 360, but is 0.08 dB lower on Tanks and Temples; LPIPS is slightly higher by 0.0008 on DL3DV and slightly lower on the other two datasets. Only SB alone with the original Gaussian primitives outperforms PBE on the Temple dataset in terms of PSNR and LPIPS; in most cases, the PBE method is optimal. In addition, after removing the fusion of multi-view features (not the multi-view in the geometric basis), all metrics decreased somewhat: PSNR drops by 0.02–0.24, while SSIM and LPIPS change only slightly, showing an overall decline. Together with Table 5, this shows that multi-view is effective in both training and inference.