acceptodds
Under review as a conference paper at ICLR 2027

LoCoSplat: Real-Time Feed-Forward 3D Gaussian Splatting with Minimal 3D Reasoning

10 views

Abstract

Feed-forward 3D Gaussian Splatting (3DGS) increasingly aggregates multi-view evidence with heavy learned 3D networks. We propose LoCoSplat (Local-Context Splatting), motivated by the observation that a Gaussian is a local primitive: once depth is predicted, what the 3D stage must add (scale, rotation, opacity) depends on the point cloud around each anchor, and a fixed local average of that neighbourhood is enough to supply it, no heavy network required. LoCoSplat realises exactly this average: it splats a 16-d linear projection of the point features into a fine and a coarse grid and reads both back at each anchor with a 0.14M-parameter pointwise MLP; with no learned 3D network and no dynamic sparse computation, its whole encoder runs as one fp16 CUDA graph. On RealEstate10K, LoCoSplat outperforms every prior feed-forward method on PSNR, SSIM, and LPIPS at 6, 12, and 24 views, with a margin that widens as views densify (+3.3 PSNR over VolSplat, the prior voxel-aligned state of the art, at 24 views) and grows further under zero-shot transfer to ACID and fine-tuning on ScanNet. It reconstructs a 6-view scene in 33 ms on one NVIDIA RTX PRO 6000 GPU, the fastest of seven feed-forward methods and 4.2× faster than the previous state of the art, trains 2.7× faster (5.9× at 24 views), and uses 6.7× less inference memory.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, feed-forward 3DGS, real-time, novel view synthesis, multi-view learning, efficient neural rendering, local context aggregation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_CgwfOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This work presents LoCoSplat, a feed-forward 3D Gaussian Splatting framework that replaces a learned sparse 3D U-Net with a fixed two-scale grid average and small pointwise networks. The central claim is that, after depth prediction and voxelization, each Gaussian mainly needs low-order local geometric context rather than a high-capacity learned 3D module. On RealEstate10K, the method reports the best PSNR, SSIM, and LPIPS among the evaluated feed-forward methods at 6, 12, and 24 input views, while reconstructing a six-view scene in 33 ms and using substantially less memory than VolSplat. The paper also reports zero-shot ACID and fine-tuned ScanNet results, controlled ablations of the 3D stage, and an analysis attributing dense-view degradation to the plane-sweep candidate spacing in the depth branch rather than to 3D aggregation.

Critical strengths
  1. Main design of proposed method is well motivated and practically important. Replacing dynamic sparse 3D computation with fixed grid averages and pointwise maps makes the encoder compatible with a single static CUDA graph, and the reported latency and memory gains are substantial.

  2. Empirically, LoCoSplat performs strongly along several dimensions. It leads every evaluated image-quality metric at 6, 12, and 24 views on RealEstate10K, surpasses a retrained VolSplat baseline, and stays competitive or better when transferred zero-shot to ACID or fine-tuned on ScanNet.

  3. Architectural and systems gains are disentangled effectively in the cumulative ablations. Removing the U-Net by itself improves both reconstruction quality and speed, after which the deployment stack cuts latency further with essentially no change in the reported metrics.

Critical weaknesses
  1. Generality remains established only within a fairly narrow operating regime. The experiments use fixed voxel sizes of 0.001 and 0.128, posed inputs, and 256×256 images. As the paper itself notes, those scales may not carry over to 512×512 resolution, wider baselines, noisy poses, or poorer depth, all of which matter for deploying a fixed spatial aggregation rule.

  2. Isolation of the proposed 3D mechanism is clearest only for 6 views RealEstate10K setting, primarily against VolSplat's U-Net. The headline end-to-end model simultaneously changes the plane-sweep candidate count, backbone interface, regressor width, prediction head, and deployment path. Consequently, the full quality and runtime improvements cannot be assigned solely to two-grid aggregation.

  3. Qualitative examples are selected through a procedure that favors LoCoSplat: frames are removed when any baseline has higher PSNR or when LoCoSplat falls below 26 dB, and the remaining cases are ranked for inspection of local failures. The authors disclose this process, but the figures therefore emphasize selected successes rather than typical outputs or a balanced set of failure cases.

Questions for authors
  1. How do reconstruction quality and memory respond to the fine and coarse voxel sizes as image resolution, focal length, or scene scale changes? Could the authors give a normalized rule for setting these values, or report experiments at 512×512 and on data with wider baselines?

  2. For 12 and 24 views, what portion of the gain over retrained VolSplat comes from the two-grid 3D stage itself, as opposed to reducing the plane-sweep candidates from 128 to 64 and modifying the backbone? A factorial study at these view counts would make the attribution clearer.

  3. Could the evaluation include randomly sampled visual results and representative cases in which a competing method outperforms LoCoSplat, alongside the existing diagnostic selection? Such examples would provide a more interpretable visual comparison.

  4. Is the depth collapse at 24 views consistent across training seeds, and can different optimization or regularization prevent it while preserving 128 candidates? The current experiments connect the failure to disparity per candidate, but do not fully explain why the learned output becomes an input-independent one-hot distribution.

1 comment

We thank the reviewer for recognizing our design, results and ablations, and for the constructive questions and suggestions. We have uploaded a revised paper with all changes marked in blue, and respond to each point below.

W1 (noisy poses, poorer depth). We added noise to the input poses (0.5-2∘2^\circ rotation, 1-5% translation) and to the predicted depth (3-5%) on all 709 RealEstate10K test scenes, without retraining. LoCoSplat stays ahead of VolSplat at every level (+0.25 to +0.61 PSNR); pose errors hurt both models equally, so the learned U-Net does not handle them better than our fixed average. Details are in Sec. 3.5 and App. C.6 (Tab. 15).

ConditionLoCoSplatVolSplatΔ\Delta PSNRLoCoSplat better
Clean31.4930.95+0.5475%
Pose 0.5∘0.5^\circ / 1%26.3825.91+0.4771%
Pose 1∘1^\circ / 2%23.6223.23+0.3968%
Pose 2∘2^\circ / 5%20.8620.61+0.2563%
Depth per pixel 3% / 5%31.32 / 30.9530.74 / 30.34+0.58 / +0.6176% / 79%
Depth per view 3%31.2430.65+0.6075%

W1 / Q1 (voxel sizes vs. resolution, focal length, scene scale; 5122512^2). vv and HH need no retuning: the rule is vv sub-pixel (so anchors are pixels) and H≈4%H \approx 4\% of the median depth, which 0.0010.001/0.1280.128 satisfy. Across scene scales, the gain over VolSplat does not vary with HH (App. C.4; see our response to Reviewer ttZu, Q4). Without retraining, from 2562256^2 up to the native 352×640352\times640 and at 1.25×1.25\times/1.5×1.5\times focal length, vv stays ≤0.12\le0.12 px and HH at 3.5-4.7% of the depth, and rescaling HH at test time by 0.67-1.375×1.375\times changes PSNR by ≤0.13\le0.13 dB. Memory and time grow roughly linearly with pixel count: at every size up to 512×896512\times896, LoCoSplat's encoder is 1.91.9-2.4×2.4\times faster than VolSplat's with 1.9×1.9\times less memory, and our deployed path needs only 153 ms / 3.4 GB at 5122512^2 (8×8\times less memory than VolSplat). Details are in App. C.7 (Tab. 16).

W2 / Q2 (attribution at 12 and 24 views). As suggested, we added the factorial study (App. C.3, Tab. 10). The two-grid model changes only VolSplat's 3D stage, keeping its backbone and D=128D{=}128. At 12 views, the 3D stage alone gives +0.58 of LoCoSplat's +0.61 PSNR. At 24 views, DD decides the collapse, and the two-grid 3D stage stays ahead of the U-Net at either DD (+0.70 at D=128D{=}128, +0.17 at D=64D{=}64). The speedup also comes from the 3D stage: at equal D=64D{=}64 our encoder is still 3.7×3.7\times faster.

ViewsModelDDPSNR ↑\uparrowPaired Δ\Delta PSNR vs. VolSplat at the same DD
12VolSplat12830.86---
12Two-grid model12831.43+0.58
12LoCoSplat6431.47+0.61 (vs. D=128D{=}128)
24VolSplat12827.24---
24Two-grid model12827.94+0.70
24VolSplat6430.40---
24LoCoSplat6430.57+0.17

W3 / Q3 (selected qualitative examples). As suggested, App. E.5 now adds randomly sampled results and failure cases where a baseline beats LoCoSplat (Figs. 14-15), plus per-baseline win rates over all test scenes (Tab. 21).

Q4 (collapse across seeds; prevention with 128 candidates). We retrained LoCoSplat with D=128D{=}128 at 24 views with two new seeds: one collapses and one does not. In total, 4 of 5 D=128D{=}128 trainings collapse, while every D=64D{=}64 model is healthy. The collapse is therefore an optimization failure, decided within the first few hundred steps, when a collapsing run's matching distribution locks into a one-hot it never leaves. It can be prevented with all 128 candidates kept: matching each view against its second-nearest cameras restores the disparity per candidate. On the collapsing seed, this keeps the model healthy, as we predicted before training, and gains +5.1 PSNR (25.37 to 30.48). Details are in Sec. 3.5 and App. D.4 (Tab. 20).

Reviewer_ttZuOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes LoCoSplat, a real-time feed-forward 3DGS model that builds on the depth-guided voxel-aligned pipeline of VolSplat without its sparse 3D U-Net. In particular, anchor features are averaged in a fine and a coarse grid, fused by a 0.14M-parameter MLP and decoded by a tiny pointwise head. This allows the whole encoder to run as one CUDA graph. Experiments on RealEstate10K at 6, 12 and 24 views, with transfer to ACID and ScanNet, report improvements over VolSplat with 4.2× faster reconstruction.

Critical strengths
  • The idea of questioning how much learned 3D capacity a voxel-aligned pipeline really needs is interesting and well motivated. Ablation studies are carefully conducted with paired per-scene comparisons and repeated seeds, which makes the finding credible at the tested voxel size.

  • The efficiency gains are substantial, with clearly lower runtime and peak memory than VolSplat, making LoCoSplat a practical baseline for real-time feed-forward 3DGS.

Critical weaknesses
  • The dense-view margin over VolSplat seems over-claimed. The abstract highlights "+3.3 PSNR over VolSplat, the prior voxel-aligned state of the art, at 24 views" (L024-025). However, App. D.1 traces this gap to the plane-sweep candidate count, a failure that LoCoSplat "avoided incidentally through a change made for speed" (L1117). VolSplat retrained with D=64 only appears in Fig. 6 as a partial training curve with PSNR at 25k and 50k steps. The fair full-budget 24-view comparison is thus missing. I suggest reporting the full-budget 24-view VolSplat with D=64 in Tab. 1 and toning down the dense-view claims accordingly.

  • Limited evidence for the minimal 3D reasoning claim. The finding is only shown in an effectively pixel-aligned regime. At the fine voxel size of 0.001, 99.6% of anchors have no occupied neighbor (L073). Even the retrained VolSplat keeps 393,175 Gaussians at 6 views in Tab. 7, nearly the 393,216 input pixels. In this regime the U-Net mainly acts as a learned smoothing over its receptive field (App. C.2). It seems that a fixed coarse average plays a similar role. I am not convinced this supports the claim that "this 3D stage needs very little" (L471). It also seems inconsistent with the motivation of escaping "resolution-bound Gaussian budgets" (L038). It is highly recommended to repeat the comparison of Tab. 9 at coarser voxel sizes, where each anchor merges many pixels.

Questions for authors
  1. How fast does VolSplat run with D=64? It would be better to report how much of the 4.2× speedup remains at equal D.

  2. Since scale, rotation and opacity now come from a coarse average, I wonder how well the Gaussians align with surfaces. It might be better to visualize depth maps and Gaussian point clouds, and to provide video renderings.

  3. The qualitative figures discard frames "where any baseline beats LoCoSplat in PSNR" (L1171-1172). It might be better to show randomly sampled examples and a few failure cases for a fair comparison.

  4. Are v=0.001 and H=0.128 kept unchanged when fine-tuning on ScanNet? How sensitive are the results to the scene scale of each dataset?

1 comment

We thank the reviewer for recognizing our motivation, ablations and efficiency, and for the constructive questions and suggestions. We have uploaded a revised paper with all changes marked in blue, and respond to each point below.

W1 / Q1 (24-view margin and speed at equal DD). As suggested, we retrained the original VolSplat at 24 views for the full 150k steps with D=64D{=}64 as the only change. It no longer collapses and gains +3.16 PSNR over D=128D{=}128, confirming that most of the 24-view gap comes from the candidate count. LoCoSplat stays ahead of it on every metric (+0.17 PSNR paired, better in 55% of scenes). On speed, D=64D{=}64 makes VolSplat only 4-8% faster, so at equal DD our encoder is still 3.7×3.7\times faster on the deployed path and 2.22.2-2.5×2.5\times faster in fp32 at 6 to 24 views: the speedup comes from the 3D stage. We added this baseline to Tabs. 1 and 5 and the equal-DD timings to Sec. 3.3, and toned down the dense-view claims in the Abstract, Introduction and Sec. 3.2.

24 viewsDDPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowEncoder (ms)Peak mem. (GB)
VolSplat (retrained)12827.240.8810.12754227.3
VolSplat (retrained)6430.400.9320.08350220.7
LoCoSplat, fp326430.570.9340.08220414.2
LoCoSplat, deployed (fp16 CUDA graph)6430.570.9340.0821352.9

W2 (coarser voxel sizes). As suggested, we repeated the comparison of Tab. 9 (sparse 3D U-Net vs. our two fixed grid averages) at a coarser fine voxel size, v=0.016v{=}0.016 instead of 0.001 (H=0.128H{=}0.128 unchanged), training both models from scratch for the full 150k steps. Here 81% of anchors have an occupied 333^3 neighbour (vs. 0.4% at v=0.001v{=}0.001) and there are 22% fewer Gaussians than pixels, so the U-Net's convolutions do see neighbours. The two-grid model still wins on every metric (+0.40 PSNR paired, better in 75% of scenes), with 47×47\times fewer 3D-stage parameters. Both models keep VolSplat's head here, since LoCoSplat's pointwise head is designed for sub-pixel voxels; we note this scope in the Limitations. We also removed "resolution-bound Gaussian budgets" from the Introduction: our motivation is the cost of the learned 3D stage, not the Gaussian budget. Details are in Sec. 3.5 and App. C.8 (Tab. 17).

v=0.016v{=}0.016PSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrow3D-stage params
VolSplat (sparse 3D U-Net)30.670.9320.08617.3M
Two-grid model (no U-Net)31.070.9370.0790.37M

Q2 (surface alignment). We visualize rendered depth maps, Gaussian point clouds and Gaussian orientations on ScanNet against the sensor depth (Fig. 6, App. C.5); video renderings are available at this anonymous link. The Gaussians are sub-pixel (≈0.4\approx0.4 px) and nearly isotropic, so their orientation barely matters and their placement follows the predicted depth, which is accurate (AbsRel 0.071, δ<1.25\delta<1.25 0.96 over 72 sequences). Compared with VolSplat, our depth is more accurate (AbsRel 0.071 vs. 0.079), our Gaussian centres lie 10.6 mm closer to the surface, and orientation is no worse (Tab. 14).

Q3 (random samples and failure cases). As suggested, App. E.5 now shows randomly sampled results and failure cases at 6 and 24 views with the cause of each failure (Figs. 14-15), and per-baseline win rates over all test scenes (Tab. 21): LoCoSplat has the higher PSNR in 75% / 68% of scenes against VolSplat at 6 / 12 views, and in 74-98% against AnySplat and WorldMirror at every view count.

Q4 (v/H on ScanNet; scene-scale sensitivity). Yes, v=0.001v{=}0.001 and H=0.128H{=}0.128 are unchanged on all datasets; only near/far changes, to 0.5/15. Sensitivity to scene scale is low. vv is sub-pixel everywhere (0.07-0.17 px), so anchors are simply pixels, and the per-scene gain over VolSplat does not vary with HH's size in pixels across 1,822 ACID scenes (2.7-21 px, ρ=−0.09\rho=-0.09) or 72 ScanNet sequences (ρ=+0.11\rho=+0.11). It is in fact largest on ScanNet (+0.92), where HH is relatively coarsest (22 vs. 9 px on RealEstate10K). Details are in Sec. 3.4 and App. C.4 (Tab. 13).

Reviewer_SBkFOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LoCoSplat, motivated by the observation that a Gaussian is a local primitive: once depth is predicted, what the 3D stage must add (scale, rotation, opacity) depends on the point cloud around each anchor, and a fixed local average of that neighbourhood is enough to supply it, no heavy network required.

Critical strengths
  1. From Tab. 1 and 2, it can be seen that this article has achieved good new perspective synthesis results through relatively simple structural improvements, while improving reconstruction efficiency and reducing video memory usage.
  2. The writing of the paper is relatively clear and easy to understand.
Critical weaknesses
  1. One of the author's core innovations is to demonstrate that simple local averaging can replace the more complex 3D U-Net networks in previous work. However, the author's experiment and Table 12 show that without introducing coarse scale domain sets, using only fine grid features can achieve a PSNR of 31.38 dB, while adding coarse scale local averaging only increases it to 31.51 dB. In addition, due to the sparsity of fine grids, the vast majority of anchor points cannot directly obtain information from adjacent voxels, indicating that the model still has good reconstruction performance in the absence of additional 3D neighborhood aggregation.
  2. Lack of comparison with new methods such as tokenGS and QuerySplat.
Questions for authors
  1. Can the author further explain how much of the performance improvement comes from local averaging itself, and how much comes from the training effect improvement after removing 3D U-Net?
  2. Can the author design a control experiment to compare with other lightweight and learning aggregation modules (such as local attention or CNNs), in order to better demonstrate the effectiveness of the fixed local averaging strategy and its advantages over learning-based aggregation methods?
3 comments

We thank the reviewer for recognizing the simplicity, efficiency and clarity of LoCoSplat, and for the constructive questions and suggestions. We have uploaded a revised paper with all changes marked in blue, and respond to each point below. In short, we isolate the contribution of the coarse average and show where its gain grows (App. C.9), add TokenGS and QuerySplat to all main comparisons (Tabs. 1-3, Sec. 3.2-3.4), and compare the fixed average with a learned local attention (App. C.2).

W1 / Q1 (fine-only vs. coarse average). The gain decomposes as: removing the U-Net and using only fine-grid features (the 31.38 dB setting in Tab. 12 raised in the review; "fine-only" below) +0.45, then adding the coarse-scale local average +0.13, for +0.58 in total. That fine-only already beats the 17M U-Net is part of our claim: with isolated anchors, a sparse convolution is effectively pointwise (Remark 1). The coarse average is the only grid change whose removal hurts (seed s.d. <0.01<0.01), and it is nearly free. Its gain is broad (positive in 70% of scenes, significant in every pixel subset we tested, e.g., depth edges and low-texture regions), and it grows where anchors carry less multi-view evidence: 2.5−4.2×2.5-4.2\times with three or four input views, 2.2×2.2\times on target views outside the input-camera span, and 3.3×3.3\times zero-shot on DL3DV at 5122512^2. A wider coarse scale (12 grid levels, coarsest cell 2.048 instead of 0.128) amplifies each gain. Details are in Sec. 3.5 and App. C.9 (Tab. 18).

Setting (PSNR)Fine-onlyLoCoSplat (fine + coarse)LoCoSplat −- fine-onlyWider coarse scale (12 levels) −- fine-only
RealEstate10K, 6 views31.3831.49+0.11+0.24
4 views28.6028.87+0.27 (2.5×2.5\times)+0.34
3 views25.7826.23+0.45 (4.2×4.2\times)+1.20
6 views, targets inside the input-camera span31.0331.13+0.10+0.24
6 views, targets outside the input-camera span21.8122.04+0.23 (2.2×2.2\times)+0.78
ACID, zero-shot33.1033.18+0.09+0.19
DL3DV, zero-shot, 5122512^218.0218.38+0.36 (3.3×3.3\times)+0.72

(31.49: the 31.51 checkpoint of Tab. 1, re-evaluated together with the compared models on a different machine; the 0.02 difference is within the replicate noise.)

W2 (TokenGS, QuerySplat). We have added comparisons with both methods, using their released checkpoints. On RealEstate10K, LoCoSplat outperforms both on every metric at 6 and 12 views and in SSIM and LPIPS at 24 views; only the 24-view PSNR of TokenGS is higher, by 0.15 (30.72 vs. 30.57). On zero-shot ACID, it leads TokenGS and QuerySplat by +2.15 and +4.61 PSNR. It is also faster at 6 views: 33 ms deployed and 57 ms in plain fp32, vs. 73 ms and 652 ms (both fp32). Details are in Sec. 3.2-3.4. We have revised the corresponding claims in the Abstract, Introduction, and Sec. 3 accordingly.

Method (PSNR ↑\uparrow / SSIM ↑\uparrow / LPIPS ↓\downarrow)RE10K 6 viewsRE10K 12 viewsRE10K 24 viewsACID 6 views (zero-shot)6-view time (ms)6-view peak mem. (GB)
QuerySplat28.74 / 0.890 / 0.11829.60 / 0.909 / 0.11429.98 / 0.916 / 0.11428.57 / 0.833 / 0.176652.49.8
TokenGS30.52 / 0.926 / 0.12230.77 / 0.929 / 0.11930.72 / 0.929 / 0.12131.03 / 0.900 / 0.18972.81.2
VolSplat (retrained)30.93 / 0.938 / 0.07830.86 / 0.939 / 0.07727.24 / 0.881 / 0.12732.29 / 0.924 / 0.097137.07.3
LoCoSplat31.51 / 0.942 / 0.07431.47 / 0.942 / 0.07330.57 / 0.934 / 0.08233.16 / 0.931 / 0.09133.0 (fp32: 56.9)1.1 (fp32: 3.9)

Q2 (learned local aggregation). We trained a learned local attention at full budget: each Gaussian's coarse-scale feature becomes the coarse-cell mean plus a zero-initialised query-key-value attention over the 333^3 surrounding coarse cells (4 heads, query from the Gaussian's own feature). It gains only +0.09 PSNR (31.57 vs. 31.49), but makes the encoder 1.6×1.6\times slower (81 vs. 51 ms), needs 2.3×2.3\times the memory, and cannot be captured in the single CUDA graph we deploy (62 vs. 32 ms deployed). The fixed average thus gives nearly the same quality at a fraction of the cost. Details are in App. C.2 and the Limitations.

Coarse-scale aggregationPSNR ↑\uparrowSSIM ↑\uparrowLPIPS ↓\downarrowEnc. fp32 (ms)Enc. deployed (ms)Peak mem. fp32 (GB)
Fixed local average (LoCoSplat)31.490.9420.07451.432.0 (full CUDA graph)3.9
Average + learned local attention31.570.9420.07381.1 (1.6×1.6\times)61.8 (3D stage not graph-capturable)8.8 (2.3×2.3\times)
Reviewer_SBkFOpenReview ↗

Response to Authors' Clarifications

Thank you for your comprehensive response. All of my concerns have been addressed, and I will therefore raise my score to Weak Accept.

We are glad to have addressed your concerns. Thank you for your constructive suggestions, which have helped improve the paper, and for raising your score.

Comments on the paper

3 comments
AuthorsOpenReview ↗

General Response

We thank all reviewers for their constructive feedback. We have uploaded a revised paper with all changes marked in blue, and respond to every point in the individual replies. The main changes are:

  • TokenGS and QuerySplat added as baselines (Sec. 3.2-3.4, Tabs. 1-3).
  • VolSplat retrained at 24 views with D=64D{=}64; dense-view and speed claims revised (Abstract, Sec. 1, 3.2, 3.3, Tabs. 1, 5).
  • U-Net vs. fixed average at a coarser voxel size (Sec. 3.5, App. C.8).
  • Attribution of the 12- and 24-view gains (Sec. 3.5, App. C.3).
  • Comparison with a learned local attention (Sec. 3.5, App. C.2).
  • Where the coarse average helps most (Sec. 3.5, App. C.9).
  • Robustness to noisy poses and depth (Sec. 3.5, App. C.6).
  • Input resolution, focal length and scene scale (Sec. 3.3, 3.4, App. C.4, C.7).
  • Depth and Gaussian geometry on ScanNet (Sec. 3.4, Fig. 6, App. C.5), with video renderings at this anonymous link.
  • Random samples, failure cases and win rates (Sec. 3.2, App. E.5).
  • Collapse across seeds and its prevention with 128 candidates (Sec. 3.5, App. D.4).
  • Expanded Limitations (Sec. 5).

Our core conclusions are unchanged and now hold beyond the originally tested setting.

AuthorsOpenReview ↗

Follow-up on our rebuttal

Dear Reviewers,

We submitted our rebuttal several days ago and would appreciate hearing whether you have any remaining questions or concerns. We would be happy to address any unresolved points and provide further clarification.

Thank you again for your time and feedback!

The Authors

AuthorsOpenReview ↗

Rebuttal Summary

Dear AC and Reviewers,

We sincerely thank you for your constructive feedback, which has substantially improved the paper. We briefly summarize the discussion below.

Main Strengths. We are pleased that all three reviewers are now positive about the paper, and that Reviewer SBkF raised their score after we addressed their concerns. All reviewers recognize the simple, well-motivated design of LoCoSplat and its substantial gains in runtime and memory, which make it a practical real-time baseline. Reviewers also highlighted:

  • Strong novel view synthesis results (SBkF), across view counts and datasets (Cgwf).
  • Careful ablations, with paired per-scene comparisons (ttZu), that effectively separate architectural and systems gains (Cgwf).

Main Revisions (marked in blue in the revised paper):

  • Core claim: We retrained both models at a coarser voxel size, where the fixed average still beats the 3D U-Net (ttZu). We also trained a learned local attention, which the fixed average nearly matches at a fraction of the cost. New evaluations show that its gain grows where multi-view evidence is weak (SBkF). A new factorial study isolates the 3D stage's contribution at 12 and 24 views (Cgwf).
  • Comparisons: We added TokenGS and QuerySplat as baselines (SBkF). We retrained VolSplat at 24 views with D=64D{=}64: it still trails LoCoSplat, which remains 3.7×3.7\times faster at equal DD. We toned down the dense-view claims accordingly (ttZu).
  • Generality: New experiments show that LoCoSplat keeps its lead over VolSplat under noisy poses and depth. They also show that the same voxel sizes work across resolutions, focal lengths and scene scales (ttZu, Cgwf). New retrainings across seeds show that the 24-view collapse is an optimization failure, which can be prevented while keeping all 128 candidates (Cgwf).
  • Transparency: We added ScanNet depth and Gaussian geometry, video renderings, random samples, failure cases and win rates (ttZu, Cgwf), and expanded the Limitations.

Thank you again for your time and effort.

Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.