- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper proposes Bounded Appearance Field (BAF) for in-the-wild 3D Gaussian Splatting. It combines a fixed-capacity 3D appearance field with global and per-Gaussian color corrections. Reference residuals are projected into the field for appearance adaptation, while DINOv3-based uncertainty weighting helps suppress transient objects.
Critical strengths
-
The idea of decoupling shared appearance capacity from Gaussian densification is reasonable and practically useful.
-
The method achieves competitive results on Photo Tourism and NeRF On-the-go with only 30k training steps.
-
The paper is well organized, and the supplementary material provides useful implementation details and ablations.
Critical weaknesses
-
The novelty is somewhat limited. Shared appearance fields and reference-guided appearance modeling have been explored in SWAG, Wild-GS, and WilLaGS. The DINO-based uncertainty strategy is also close to WildGaussians. The main contribution seems more like a combination of existing techniques.
-
The ablations do not strongly support the core contribution. In Table 3, adding the shared field without reference conditioning even hurts PSNR. The full model improves Photo Tourism PSNR by only 0.19 dB over the global/local-only baseline. It is unclear how much the proposed field really contributes.
-
Some important comparisons are missing. There are no direct results against SWAG, Wild-GS, or WilLaGS. Also, a comparison with simpler grid or triplane representations is needed to justify using Mamba3.
-
The efficiency comparison is not fully fair. The 30k WildGaussians baseline still uses its original 200k training schedule, and some memory results come from different GPUs. The proposed method also takes longer than the local 30k WildGaussians run.
-
The experimental validation is still incomplete. The ablations use single runs, and there is limited analysis of different reference sizes and Gaussian counts. Moreover, Appendix A.5 mentions that reference sampling may access pixels outside the observed region, which raises a concern about evaluation fairness.
Questions for authors
-
Can the authors provide a stronger comparison with Wild-GS and WilLaGS, and clarify what is fundamentally new in BAF?
-
Why is Mamba3 necessary? Would a simple MLP, 3D convolution, or hash-grid field achieve similar results?
-
Can the authors provide fair 30k-step baseline comparisons with tuned schedules, matched hardware, and training-time measurements?
Response to Weaknesses and Questions, with Revised Claims, Mamba3 Controls, Protocol-Matched Comparisons and Exclusive-GPU Timing
Dear Reviewer,
Thank you very much for your careful review and constructive comments. We are encouraged that you find the idea of "decoupling shared appearance capacity from Gaussian densification" to be "reasonable and practically useful." Your comments also helped us identify where the submission overstated its claims, and we have revised the paper to address each of them.
W1, W2 (novelty and ablations). Thank you for pointing out that "the DINO-based uncertainty strategy is also close to WildGaussians." This strategy was inspired by WildGaussians, and our modifications consist of the newer DINOv3 backbone and a feature-distance confidence on NeRF On-the-go. Likewise, a fixed-capacity shared field and 3D-aligned reference conditioning have precedents in SWAG and Wild-GS, respectively. The novelty is therefore narrower than we claimed. What remains new is that we project color residuals and coverage from the observed part of a test view into a fixed-capacity grid, which enables appearance adaptation without per-image optimization. On fixed checkpoints, the reference raises PSNR by 1.46 ± 0.10 dB without the test-time embedding fit and has no measurable effect with the fit. We are also grateful for your comment that it is "unclear how much the proposed field really contributes." Accordingly, our revised claim is that the shared field decouples appearance capacity from Gaussian densification and that the full model remains competitive with the global/local-only baseline (0.20 ± 0.11 dB lower PSNR with equal SSIM and LPIPS over three seeds).
Q2 (Mamba3). A token-wise MLP and a 3D convolution indeed "achieve similar results," so we do not claim that Mamba3 is necessary. Replacing Mamba3 with these alternatives lowers PSNR on Photo Tourism (PT) by 0.20 and 0.31 dB, respectively. Both differences are within run-to-run variation, since a rerun with the same seed shifts the PT average by 0.38 dB. We retain Mamba3 because it attains the highest PSNR on PT, with the largest margin on Brandenburg Gate, and requires fewer parameters than the MLP. Hash-grid and triplane variants are left for future work.
W3, Q1 (SWAG and Wild-GS). SWAG follows the same left-half protocol, so we have added it to Table 1, where our method attains higher PSNR on all three scenes. Wild-GS, in contrast, evaluates at half resolution (R/2) and takes appearance from the full test image, so the evaluated region is also an input. We evaluated our Table 1 checkpoints, which were trained at full resolution, under this protocol. The embedding fit and the reference read the full test image, and the metrics are computed on the right half. Our method then reaches 30.97 / 26.59 / 25.48 dB PSNR on Brandenburg Gate, Sacre Coeur and Trevi Fountain, which is 1.32, 1.60 and 1.03 dB above the values reported by Wild-GS. SSIM and LPIPS are also better on every scene. Our average PSNR is also 0.65 dB higher computed on the full image. We leave Wild-GS out of Table 1 only because its protocol differs from that of the table.
W3, Q1 (WilLaGS). The arXiv and ECCV 2026 versions of WilLaGS report mutually inconsistent results, and no code has been released to reproduce them. We thus have reservations about these results and do not compare with WilLaGS.
W4, Q3 (efficiency). The efficiency comparison was indeed "not fully fair." Our 30k WildGaussians baseline was a 200k-schedule run truncated at 30k updates, and our training times were measured on a shared GPU. We reran the baseline with its own 30k schedule and repeated all timing measurements on a single RTX 5090 with exclusive access. With this schedule, WildGaussians reaches 22.23 / 0.821 / 0.234 (PSNR / SSIM / LPIPS), which is 2.5 dB below our PSNR. At an equal number of updates, our method still "takes longer" (48.8 versus 40.1 min), as it uses four images per update. Under the NerfBaselines definition, our peak GPU memory is 4.07 versus 5.24 GiB on PT, but higher on NeRF On-the-go. On the same GPU, our method reaches the SSIM and LPIPS of the released models after 20 to 43 and 23 to 48 min, respectively, whereas the 200k schedule takes 246 to 453 min. In PSNR, our runs remain 0.04 to 0.48 dB below these models. We have corrected the corresponding table and text.
W5 (validation). Reference sampling is now confined to the observed left half, which changes the PSNR of the submitted checkpoints by only 0.007 dB. The principal comparisons now use three seeds. Regarding reference size, a test-time reference covering a quarter of the image raises PT PSNR by 0.01 dB, and one covering an eighth lowers it by 0.10 dB. Regarding Gaussian counts, the final counts of the ablation models lie within 6% of each other, and the Mamba3 mixing step takes approximately 1.5 ms per image regardless of the count.
Thank you very much again for your time and valuable feedback. We would welcome any further comments.