acceptodds
Under review as a conference paper at ICLR 2027

Bounded Appearance Field for In-the-Wild 3D Gaussian Splatting

Abstract

In unconstrained photo collections, appearance variations and transient occluders violate multi-view consistency, a key assumption of 3D Gaussian Splatting (3DGS). Methods with per-Gaussian appearance features tie spatial appearance capacity to geometric densification, although finer geometry need not require more complex regional appearance. We introduce a Bounded Appearance Field (BAF) that combines a shared-field of fixed capacity with global color adjustments and per-Gaussian local corrections. A fixed-resolution token grid is mixed once per image and queried by Gaussians through trilinear interpolation. This keeps shared field capacity independent of the Gaussian count, while local storage and per-Gaussian computation remain linear in that count. To support adaptation from partial observations of a held-out view, we project color residuals and coverage into the grid, providing spatially aligned appearance evidence for predictions outside the observed region. We additionally use frozen DINOv3 feature distances to supervise a heteroscedastic 2D uncertainty head. Its predicted weights modulate the photometric loss and the gradients driving densification to reduce the influence of transient observations. On both Photo Tourism and NeRF On-the-go datasets, our method converges within 30k updates and matches or surpasses state-of-the-art results. Our code will be available after acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Novel view synthesis, 3D Gaussian splatting, unconstrained image collections, appearance modeling, transient occlusion, uncertainty weighting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_2N1HOpenReview ↗
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
  1. The idea of decoupling shared appearance capacity from Gaussian densification is reasonable and practically useful.

  2. The method achieves competitive results on Photo Tourism and NeRF On-the-go with only 30k training steps.

  3. The paper is well organized, and the supplementary material provides useful implementation details and ablations.

Critical weaknesses
  1. 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.

  2. 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.

  3. 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.

  4. 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.

  5. 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
  1. Can the authors provide a stronger comparison with Wild-GS and WilLaGS, and clarify what is fundamentally new in BAF?

  2. Why is Mamba3 necessary? Would a simple MLP, 3D convolution, or hash-grid field achieve similar results?

  3. Can the authors provide fair 30k-step baseline comparisons with tuned schedules, matched hardware, and training-time measurements?

1 comment
AuthorsOpenReview ↗

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.

Reviewer_z5TrOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes BAF, which combines a fixed-capacity shared 3D appearance field with global and per-Gaussian color corrections. Color residuals and coverage from partial reference images support appearance prediction, while DINOv3-guided uncertainty weighting reduces transient interference. The method is evaluated on Photo Tourism and NeRF On-the-go.

Critical strengths

1、Clear motivation and architecture. The fixed-capacity shared field provides a 3D pathway for partial reference information while retaining global and local appearance corrections.

2、Competitive results and informative ablations. The paper evaluates the method on two benchmarks and explicitly analyzes the benefits and limitations of the shared field under different settings.

Critical weaknesses

1、The independent contribution of shared-field reference propagation remains unclear. In Table 3, adding the shared field reduces average Photo Tourism PSNR from 24.523 to 24.037 dB; adding reference input then raises it to 24.712 dB. These results do not isolate the benefit of reference residuals from that of Mamba3-based cross-token interaction. A control replacing Mamba3 with a token-wise MLP, while retaining the same grid, reference input, and global/local branches, would clarify this distinction.

2、BAF shares several design choices with WildGaussians in appearance correction and uncertainty-based training, while shared spatial representations and 3D-aligned reference conditioning have precedents in SWAG and Wild-GS, respectively. The authors are encouraged to clarify the newly introduced components and compare variants with and without cross-token mixing while retaining the same reference input, grid configuration, and training settings. Comparisons with alternative shared spatial representations of comparable capacity would also help establish the independent contribution of the proposed design.

3、The overall framework diagram could be improved for clarity. Figure 1 contains numerous modules, arrows, and connections, making the overall architecture difficult to follow. In particular, the relationships among the Bounded Shared-Field, Scene-Aligned Reference Conditioning, and Global/Local Decoders are not sufficiently clear.

4、Although Figures 2–5 provide qualitative comparisons and appearance ablations, they do not explicitly demonstrate appearance interpolation or multi-view rendering under a fixed appearance condition. Such visualizations would help assess the learned appearance representation beyond per-image reconstruction. For the reference-conditioned model, please clarify how the appearance embedding and reference grid are held fixed or varied.

Questions for authors

1、How does performance change when cross-token mixing is removed while retaining the same reference input and grid? 2、Although the paper provides appearance-adaptation comparisons, it does not explicitly show appearance interpolation or multi-view renderings under a fixed appearance condition. Could the authors provide these visualizations to assess the smoothness of appearance transitions and cross-view appearance consistency?

1 comment
AuthorsOpenReview ↗

Response to Weaknesses and Questions, with the Requested Controls, a Clarification of Figure 1 and Appearance Visualizations

Dear Reviewer,

Thank you very much for your constructive review. Your two main concerns are whether the reference makes an "independent contribution" beyond cross-token mixing and what is new relative to prior work. We address each point below and revise the paper accordingly.

W1, Q1 (cross-token mixing). We evaluate your proposed control, "replacing Mamba3 with a token-wise MLP, while retaining the same grid, reference input, and global/local branches" (seed 42). PSNR decreases by 0.20 dB on Photo Tourism (PT) and 0.05 dB on NeRF On-the-go (OTG), within run-to-run variation (a rerun with the same seed and code shifts the PT average by 0.38 dB). A four-layer 3D convolution of comparable capacity is 0.31 dB lower on PT and 0.09 dB higher on OTG. The three mixers thus perform comparably, so we no longer claim that Mamba3 is necessary. We retain Mamba3 as it attains the highest PSNR on PT, with the largest margin on Brandenburg Gate (0.52 and 1.20 dB above the other two), where the seeds of the full model vary least. It also scans the whole grid with fewer parameters than the MLP (27.4k versus 98.9k) at about 1.5 ms per image. Hash-grid and triplane variants remain future work owing to time constraints.

W1 (reference contribution). We fix the checkpoints and vary only the test-time input on PT (three seeds). Δ is the PSNR gain of the full model D with the aligned reference over each row.

VariantNo embedding fit (D 19.61 dB)128-step fit (D 24.34 dB)
No reference1.46±0.100.09±0.20
Mean-pooled residuals0.42±0.010.16±0.13
Shuffled residuals1.44±0.100.09±0.19
Trained with pooled reference0.10±0.080.28±0.11
B (no shared field)0.86±0.07−0.20±0.11

Without the fit, every Δ is positive and the reference improves 140 of 150 views. Only the 0.10 dB gap and all gaps with the fit lie within run-to-run variation. We thus claim appearance adaptation without per-image optimization, not a quality gain from spatial alignment. Thank you for pointing out that "adding the shared field reduces average Photo Tourism PSNR." We now state that the shared field decouples appearance capacity from densification, with D competitive with B (last row, equal SSIM and LPIPS).

W2 (prior work). Thank you for asking us "to clarify the newly introduced components." Our uncertainty weighting follows WildGaussians but uses DINOv3 and a feature-distance confidence on OTG, as Section 3.3 now states. WildGaussians stores appearance per Gaussian, so its capacity grows with densification. SWAG, now in Table 1, has a fixed-capacity hash grid conditioned only on a global embedding, and Wild-GS lifts the full reference image into a triplane. Neither exploits spatially aligned evidence from a partially observed view, which prevents a fair comparison under the same reference input. Our fixed-capacity grid receives residuals and coverage from only the observed half, enabling appearance adaptation without per-image optimization.

W3 (Figure 1). We are grateful for your comment that "the relationships among the Bounded Shared-Field, Scene-Aligned Reference Conditioning, and Global/Local Decoders are not sufficiently clear." The reference enters only through the Bounded Shared-Field, which corrects the color directly and through the Local Decoder, whereas the Global Decoder is independent of both. Reference Conditioning computes the mean RGB residual between observed and base colors and its coverage at each lattice site. This vector is appended to the matching token of the shared field, its only link to the model. Mamba3 mixes these tokens once per image. Each Gaussian queries the field by trilinear interpolation for a direct color correction and a context vector that enters the Local Decoder. The Global Decoder predicts an image-wide correction from the image embedding alone. The three corrections are summed into a bias and scale for the SH colors. Consistency Weighting acts only during training. In the revision, we attempt to clarify Figure 1 further, and its caption now states these relationships.

W4, Q2 (visualizations). Thank you for this suggestion. The appendix now shows appearance interpolation and fixed-appearance renderings for the three PT scenes. Two training photos per scene are chosen by a fixed rule (A closest to the mean embedding, B farthest from A). The interpolation varies the image embedding and reference grid jointly in five steps. For fixed appearance, both are held at photo A across five test views. The transitions are smooth, for example from overcast daylight to night lighting on Sacre Coeur, and the fixed appearance is consistent across views. In the evaluation protocol, the reference grid is instead built once from the observed half and fixed while the embedding is fitted.

Thank you very much again. Your suggestions make the role of each component clearer, and we would welcome any further comments.

Reviewer_kyk1OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses appearance variation and transient occlusions in scene-specific 3D Gaussian Splatting. Its appearance module combines a fixed-resolution shared 3D grid, mixed once per image with Mamba3, with global affine adjustments and per-Gaussian local corrections. For partially observed reference images, color residuals and coverage are projected into the grid to condition appearance outside the observed region. A DINOv3-supervised uncertainty head supplies weights for reconstruction losses and the gradients used in densification. The method is evaluated on three Photo Tourism scenes and six NeRF On-the-go scenes, with layered appearance ablations and resource measurements.

Critical strengths

1.The complete system reports competitive reconstruction quality. On Photo Tourism, the reported averages are 24.712 PSNR, 0.8612 SSIM, and 0.1677 LPIPS, compared with 24.650/0.8507/0.1793 for the published 200k-step WildGaussians result. On NeRF On-the-go, the method obtains the best average LPIPS among the listed methods and improves average PSNR over WildGaussians from 22.15 to 22.36. These are useful system-level results, although their interpretation is subject to the protocol concerns below.

2.The construction is coherent and the ablations are informative. The paper clearly limits the fixed-capacity claim to the shared field and acknowledges that local storage and decoding remain linear in the Gaussian count. The detached uncertainty weighting has a clear role in the reconstruction gradients. Tables 3, 7, and 8 separate global/local corrections, the shared field, and reference conditioning; the paper also reports unfavorable component results and uses training-only development splits for selected design choices. This transparency helps identify what is and is not supported by the experiments.

Critical weaknesses

1.The central shared-field/reference contribution is not convincingly established under a clean evaluation protocol. Table 3 shows that adding the shared field to the global/local model reduces Photo Tourism PSNR from 24.523 to 24.037 and NeRF On-the-go PSNR from 22.393 to 22.314. There is a small LPIPS improvement on the latter, but no consistent reference-free benefit. Adding reference input improves Photo Tourism over configuration C without reference input, yet relative to the simpler configuration B, the full system D increases average PSNR by only 0.189 dB and SSIM by 0.0013, while reducing LPIPS by 0.0009. Table 7 further shows that this benefit is scene-dependent; B remains stronger than D on Sacre Coeur. These results do not isolate an advantage of spatial propagation over simpler reference conditioning. There is no capacity-matched control using the same explicit reference features without their spatial alignment, such as pooled or spatially shuffled residuals. Appendix A.5, lines 798-803, states that reference centers are selected within the observed window but their colors are bilinearly sampled from the full image, allowing pixels across the boundary to enter the reference features. Thus the stated implementation does not strictly enforce left-half-only conditioning when the right half is scored. The direct access is near the boundary, and the paper does not quantify its effect; it would be unjustified to attribute the gains to this alone. Nevertheless, a corrected crop-confined evaluation is required before the reference-transfer result can be accepted as evidence for the central contribution. Appendix A.2 itself identifies crop-confined sampling and equal-capacity reference controls as unresolved requirements. With single-initialization ablations and modest net average gains over B, these are consequential gaps.

2.The training-efficiency conclusion is not supported by a meaningful short-budget baseline or comparable cost accounting. The main text uses the collapse of the local 30k-step WildGaussians run to argue that it is competitive only under its 200k schedule. Appendix A.6 reveals that this run retains the 200k learning-rate horizon, a 100k densification endpoint, and an uncertainty warm-up starting at 35k. Evaluation at 30k also coincides with an opacity reset. This is an interrupted long-schedule run, not a convincing test of WildGaussians under an appropriately configured short budget. Update counts also do not represent equal work: Photo Tourism accumulates four images per update, giving 120k image-loss evaluations across 30k updates. Table 12 records 133.18 minutes for the proposed method and 71.98 minutes for the local WildGaussians run, with different timing definitions. The published 200k time uses an A100 rather than the RTX 5090 used locally. These records do not support a precise training-cost reduction. Appendix A.5 further states that dividing gradients by four is not compensated in the densification threshold, which can affect Gaussian growth independently of field capacity. Matched resource measurements and component-level Gaussian counts are needed to substantiate the compactness and efficiency explanation.

Questions for authors

1.Please enforce crop-confined reference sampling and rerun the Photo Tourism evaluation, reporting the corrected full-model and B/C/D results. During embedding fitting, are the uncertainty features and SSIM computations also restricted to the permitted half before any operation that mixes neighboring or distant pixels? Please describe this explicitly; the manuscript establishes the interpolation issue but leaves these other paths unclear.

2.Under the corrected protocol, does aligned residual-and-coverage conditioning outperform an equal-capacity reference control given exactly the same observed pixels, such as pooled or spatially shuffled residuals? Please provide repeatability evidence for the principal D-versus-B/control comparison, rather than relying only on the single initialization currently reported.

3.Can the efficiency comparison include a schedule-appropriate short-budget WildGaussians run, with evaluation away from the immediate opacity reset, and common-hardware quality-versus-time measurements? Please count image evaluations as well as updates, use the same timing scope, and report Gaussian counts and peak memory for B, C, and D under matched densification settings.

4.Which appearance embedding is used for unseen NeRF On-the-go views when there is neither reference input nor test-time fitting? Does evaluation use a fixed embedding or the unmodified base SH render? This choice is necessary to reproduce the reference-free results.

1 comment
AuthorsOpenReview ↗

Response to Weaknesses and Questions, with the Corrected Protocol, Three-Seed Results and Exclusive-GPU Timing

Dear Reviewer,

Thank you very much for your careful review. Your two main concerns are whether our central contribution holds under "a clean evaluation protocol" and whether our efficiency claim rests on "comparable cost accounting." We performed every requested measurement and have revised the paper on each point.

W1, Q1 (crop-confined sampling). Thank you for pointing out that "a corrected crop-confined evaluation is required." The submitted code sampled colors from the full image, so Gaussian centers within one pixel of the center line read the first scored column. All reference reads are now confined to the observed window in training and at test time. As a check, inverting the scored half changes the reference field by at most 2.7×10−72.7\times10^{-7}, below numerical noise (4.5×10−74.5\times10^{-7}), versus up to 0.50 with the old sampling. In the same code, confined reads lower the Photo Tourism (PT) PSNR of the nine submitted checkpoints by 0.007 dB on average. Table 1 still rises from the submitted 24.712 to 24.76, mainly because the submission used older evaluation code. In the embedding fit, the target is cropped to the left half before any other operation, so the uncertainty features, the rendering, L1 and SSIM never span the scored half. We have rewritten the appendix.

Corrected B/C/D. PSNR is 24.54 ± 0.09 for B, 24.09 ± 0.22 for C and 24.33 ± 0.17 for D, with SSIM 0.860 / 0.859 / 0.861 and LPIPS 0.169 for all. D is 0.20 ± 0.11 dB and C is 0.44 ± 0.14 dB below B (0.03 ± 0.04 dB on NeRF On-the-go). We have withdrawn the claim that the full model improves average quality over B and corrected Table 3, Section 4.3 and the conclusion. The Sacre Coeur deficit you noted in Table 7 persists (0.92 dB, every seed), and we list it as a limitation.

W1, Q2 (equal-capacity controls). We built the "pooled or spatially shuffled residuals" you suggested from the same observed pixels. On fixed D checkpoints without the embedding fit, the aligned reference outperforms pooling by 0.42 ± 0.01 dB and shuffling by 1.44 ± 0.10 dB. The reference itself adds 1.46 ± 0.10 dB (higher on 140 of 150 test views, 18.15 to 19.61 dB, versus 24.34 with the fit). With the 128-step fit, none of these differences is resolved (0.16 ± 0.13, 0.09 ± 0.19, 0.09 ± 0.20). Compared with a model trained with the pooled reference, aligned conditioning gains 0.10 ± 0.08 dB without the fit and 0.28 ± 0.11 dB with it. Even a rerun with the same seed and code shifts the PT average by 0.38 dB, so these margins lie within run-to-run variation. We no longer claim that spatial propagation improves quality over simpler conditioning, and claim only what the fixed checkpoints show, namely appearance adaptation without per-image optimization. B, C, D and the pooled control now have three seeds each.

W2, Q3 (efficiency). We appreciate your remark that our 30k WildGaussians run was "an interrupted long-schedule run." We repeated all timing measurements on one RTX 5090 with exclusive access. With its own 30k schedule (densification until 15k, last opacity reset at 12k), WildGaussians reaches 22.23 / 0.821 / 0.234, 2.5 dB below ours in Table 1. Update counts indeed "do not represent equal work." Using four images per update, we process 120k images against 30k. Our runs take 38 to 62 min and reach the SSIM of the released 200k models after 20 to 43 min and their LPIPS after 23 to 48 min. The 200k schedule takes 246 to 453 min and never reaches our SSIM or LPIPS. In PSNR, our runs remain 0.04 to 0.48 dB below the released models. Under the NerfBaselines definition, peak memory on PT is 4.07 versus 5.24 GiB (the submitted 2.04 GiB was PyTorch-allocated). On NeRF On-the-go (OTG), training time is equal and we use more memory (6.14 versus 3.59 GiB), so we limit the memory claim to PT. B, C and D share the uncompensated densification threshold, differ by at most 6% in Gaussian count and have nearly equal peak memory, so we do not claim that the shared field reduces the Gaussian count. Mamba3 mixing takes about 1.5 ms per image regardless of the count, whereas field queries scale linearly. We have added these measurements to the paper.

Q4. OTG test views use the zero image embedding decoded through the full appearance module rather than the base SH render. We have added this to the experimental setup.

Updated results. Table 1 now reports 24.76 / 0.861 / 0.168 on PT, above WildGaussians 200k in average PSNR (24.65) and better in SSIM and LPIPS on every scene. On OTG, we obtain 22.36 / 0.737 / 0.161, the best average LPIPS and second-best average PSNR, both better than WildGaussians (22.15 / 0.166), whose SSIM is higher (0.746). The appendix now lists every seed. D in Table 3 was retrained in the code base of B and C, and the appendix compares it with the Table 1 runs.

Thank you very much again. Your review has made the work more valuable, and we would be grateful for any further comments.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.