acceptodds
Under review as a conference paper at ICLR 2027

AsySplat: Efficient Asymmetric 3D Gaussian Splatting for Wide-Coverage Scene Modeling

Abstract

Recent generalizable 3D Gaussian Splatting (3DGS) models have advanced wide-coverage novel view synthesis (NVS), but at the cost of substantial redundant computation. We identify that the redundancy can be mitigated based on two observations: (i) high-precision geometry is not strictly required for high-quality NVS; (ii) appearance learning is generally easier than geometry recovery. Motivated by these insights, we propose an asymmetric architecture that decouples geometry and appearance modeling. The geometry branch processes coarse-grained tokens with most of the parameters for multi-view reconstruction, while the appearance branch operates on fine-grained tokens to capture details using significantly fewer parameters. The two branches interact through bilateral connections, enabling mutual guidance for their respective tasks. This task-aware asymmetry reduces the computational redundancy and allocates the computation more judiciously, thereby increasing parameter efficiency and enabling smaller models to achieve strong performance. Extensive controlled experiments demonstrate that AsySplat achieves strong performance with markedly fewer parameters and reduced training and inference overhead. Combined with recent advances in generalizable 3DGS, our asymmetric design performs on par with LongLRM++ using less than 10% of its reconstruction FLOPs, and with MVP using less than 50% of its parameters and 25% of its Gaussians. Code and models will be released.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Scene Modeling, 3D Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wETROpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AsySplat, an asymmetric architecture for efficient feed-forward 3D Gaussian reconstruction from wide-coverage multi-view images. The model processes coarse-grained tokens with a inter-view high-capacity geometry branch, while a intra-view lightweight appearance branch processes fine-grained tokens. Experiments demonstrate competitive novel-view synthesis quality and more impressive efficiency than LongLRM, alongside comparisons with recent methods and architectural ablations.

Critical strengths
  1. The asymmetric architecture offers a practical approach to reducing computation by combining coarse-grained cross-view processing with lightweight fine-grained appearance modeling.
  2. The insight is quite interesting, showing us that high-precision geometry is not a must for high-quality rendering and high-precision are generally easier to model than geometry recovery.
Critical weaknesses
  1. The claimed geometry–appearance specialization is not sufficiently established. Both branches exchange features and jointly predict Gaussian attributes after fusion, without direct evidence of distinct functional roles.
  2. The source of the rendering-quality improvements remains unclear. The asymmetric structure directly motivates computational savings, but does not by itself explain improved reconstruction quality. The gains may instead come from accompanying engineering choices, such as increased attention allocation.
Questions for authors
  1. What evidence shows that the two branches specialize in geometry and appearance, respectively, given that Gaussian attributes are predicted jointly from fused features?
  2. How does the current evidence distinguish “appearance is easier to model” from “additional cross-view processing is unnecessary after receiving geometry-branch features”?
  3. Under matched parameter budgets and Mamba-to-attention ratios, how much rendering-quality improvement comes specifically from the asymmetric token allocation?
  4. For table 2, does the advantage over LongLRM persist with longer post-optimization budgets, or do the quality–runtime curves converge or cross?
5 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your constructive comments and suggestions. Below, we provide clarifications and additional experimental results, responding to the weaknesses (W) and the questions (Q) in your review.


R3-W1 / Q1: Evidence that the branches specialize.

We test this directly on the trained 98M model by replacing one branch's output with its scene mean at inference on 20 DL3DV-140 scenes. Specifically, we replace every output token of one branch with its average over all views and positions.

PSNR↑δ<1.25↑
AsySplat25.080.794
appearance branch → mean13.960.731
geometry branch → mean12.620.287

The table above shows that replacing the appearance branch output with its mean severely degrades rendering, while largely preserving geometry. In contrast, replacing the geometry branch output with its mean destroys both rendering and geometry. Geometry is thus carried by the coarse geometry branch, while the fine appearance branch supplies the appearance details required for accurate rendering.

R3-Q2: "Appearance is easier" vs "cross-view processing is unnecessary after receiving geometry features".

Ablation experiments in Tab.5a answer this by removing the coarse-to-fine link and fine-to-coarse link, respectively:

PSNR↑
AsySplat19.24
w/o coarse-to-fine: appearance receives no geometry features19.10
w/o fine-to-coarse: geometry receives no appearance features17.33

If intra-frame processing were sufficient only because of the geometry features, the variant without coarse-to-fine would be expected to fail, but it loses only 0.14 dB. This directly demonstrates that appearance can be modeled from per-frame cues alone, without geometry features or cross-view interaction. Removing the opposite link costs 1.9 dB. Thus, the geometry branch relies heavily on appearance features, while appearance is easier to model and benefits little from cross-view interaction.

R3-Q3 / W2: Improvement from asymmetric token allocation.

Under matched parameter budgets and Mamba-to-attention ratios, [R1-W1/Q1] shows that asymmetric token allocation does not improve rendering, i.e., Two branches (8,8) versus asymmetric (16,8), but brings significant efficiency improvement.

To isolate the Mamba-to-attention ratio, we compare LongLRM models trained with exactly the same recipe and losses. The table below shows that, shrinking LongLRM from 142M to 99M at its 7:1 ratio costs 0.39 dB. Raising the attention ratio of the 99M model to 4:4 gains 0.52 dB, more than the parameters removed. A higher attention ratio therefore improves quality, and our model benefits from it. We also train our model with LongLRM's losses. With the losses, parameters and ratio all matched with LongLRM, our model is still better.

256×256#ParamMamba:attn ratioPSNR↑SSIM↑LPIPS↓
LongLRM142M7:118.780.5010.517
LongLRM99M7:118.390.480.530
LongLRM102M4:418.910.5100.513
AsySplat (w/ LongLRM losses)98M4:419.100.5170.494

Our improvement over LongLRM thus mainly comes from two parts. (i) The asymmetric allocation makes a high attention ratio affordable, since our inter-frame layers operate on 4× fewer tokens. (ii) Decoupled modeling keeps fine-grained tokens throughout the network at a low cost, whereas LongLRM merges its tokens to a coarser granularity at an intermediate layer.

R3-Q4: Longer optimization.

Please refer to Fig.A2 in the revision. The quality-runtime curves neither converge nor cross up to 200 iterations.

AuthorsOpenReview ↗

Thank You and a Gentle Follow-up

Dear Reviewer wETR,

Thank you again for your constructive comments. In response to your questions and suggestions, We have added experiments and clarifications to provide a more detailed breakdown of the model’s internal architecture. We hope these additions help address your concerns.

With less than two days remaining in the discussion period, we would like to kindly follow up to see whether you have any remaining concerns or would find further clarification helpful. We would be happy to provide additional explanations. Thank you sincerely for your time and thoughtful feedback.

Best regards,

The Authors

Thank you for the detailed responses and additional experiments. The controlled comparisons clarify the sources of the efficiency and rendering-quality gains. However, I remain concerned that the architectural design and current evidence do not sufficiently establish the claimed geometry–appearance specialization, beyond a division between coarse cross-view processing and fine intra-view processing. Both branches exchange information and jointly predict Gaussian attributes after fusion. Moreover, the strongly asymmetric dependence suggested by the low-resolution connection ablations in Table 5a is not reproduced in the high-resolution results in Table A3d. While I appreciate the practical efficiency gains, the central interpretation of the two branches remains insufficiently supported. I will therefore maintain my rating of Weak Reject.

Reviewer_B4tqOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes AsySplat, an asymmetric architecture for generalizable 3DGS from wide-coverage, high-resolution inputs. In particular, a heavy geometry branch processes coarse-grained tokens across all views, while a light appearance branch handles fine-grained tokens within each frame. The two branches exchange information through bilateral connections, and a grid-based sparse attention reduces the training cost. Experiments with 32 views at 960P demonstrate the efficiency of AsySplat over LongLRM, and its 110M variant claims parity with LongLRM++ and MVP.

Critical strengths
  • The idea of allocating computation according to the different needs of geometry and appearance is reasonable. This motivation is meaningful, since token count is the main bottleneck for 32 views at 960P.

  • Experiments against LongLRM under the same training recipe verify the efficiency of the asymmetric design. They show savings in parameters, FLOPs and training time, together with consistent zero-shot gains.

Critical weaknesses
  • The key observation that coarse geometry is sufficient is not well supported. The whole design rests on the claim that "high-precision geometry is not strictly required for high-quality NVS" (L013). However, the only quantitative evidence is Tab. 5c, trained at 256×256 for 10k iterations, where the variants of each backbone lie within 0.3 dB PSNR. In this setting, the asymmetric model also shows considerably worse geometry than the two-branch variant (δ<1.25 of 0.41 vs. 0.49). I am not convinced that this geometry drop remains harmless at 960P. Yet the final models provide no depth metrics beyond the qualitative depth maps in Fig. 5. In addition, test views are every 8th frame between 32 well-spread inputs (L758-761), so interpolated views may hide geometric errors. It would be better to report depth metrics and visualize the Gaussian point clouds on views further from the inputs.

  • Efficiency comparisons with recent methods are not consistent. The abstract claims parity with LongLRM++ "using less than 10% of its reconstruction FLOPs" (L025-026). Yet this number is estimated from a reimplementation, since "the code for LongLRM++ is unavailable" (L371). AsySplat-110 also reports 56 TFLOPs in Tab. A2, the same as the 98M model, even after adding the LongLRM++ decoder. Against MVP, the comparison is instead based on parameters and Gaussians, although MVP uses fewer FLOPs (53 vs. 56). Besides, AsySplat-110 adopts a different recipe with mixed-view training and no depth loss (L735-740, L790-791). Hence, against LongLRM++ and MVP, the gain of the asymmetric backbone itself is not isolated. I suggest a unified table of FLOPs including the decoder, latency and peak memory for all methods, measured under one protocol.

Questions for authors
  1. The teaser on the supplementary project page shows 20 vs. 16 TFLOPs for LongLRM and AsySplat. The same page lists 26.13 PSNR for AsySplat-110 at 32 views. The paper instead reports 131 vs. 57 TFLOPs (Tab. 3) and 26.40 PSNR (Tab. 4). Which numbers are correct? This directly affects the claimed efficiency gain over LongLRM.

  2. How does the asymmetric design differ from iLRM, which also decouples a compact scene representation from high-resolution image tokens? Beyond the reported numbers, it would be better to discuss the key differences and, if applicable, compare under the same training recipe.

  3. All zero-shot videos compare only with LongLRM. Zero-shot video comparisons with MVP, which is stronger on MipNeRF-360 in Tab. A2, would be more informative.

3 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your constructive comments and suggestions. Below, we provide clarifications and additional experimental results, responding to the weaknesses (W) and the questions (Q) in your review.


R2-W1-1: Ablations at 960P. Depth metrics of final models.

Please refer to [R1-W1/Q1] for 960P ablations. Depth metrics of final models are provided below.

DL3DV-140StandardExtrapolation
MethodPSNR↑δ<1.25↑PSNR↑δ<1.25↑
LongLRM24.100.74119.420.671
iLRM24.300.71720.130.672
MVP25.960.80419.100.608
AsySplat-11026.400.80121.820.758

Note that we do not claim that geometry is unimportant. Our claim is that, once geometry is reasonably accurate, small further refinements in its precision bring little rendering gain. The results under the standard protocol are consistent with this, e.g., MVP and LongLRM have slightly better geometry than AsySplat-110 and iLRM, respectively, yet both have lower PSNR. A cross-method comparison also differs in architecture and training, so the controlled test of this claim is given in [R1-W1/Q1].

R2-W1-2: Extrapolation results.

For each scene, we split the trajectory at 75% of its frames. The 32 input views are selected by k-means only from the first 75%, and the test views are every 8th frame of the last 25%. The table above shows that our method performs best under this setting in both rendering and geometry. This suggests that our method yields more globally consistent geometry, which transfers better to views far from the inputs. The Gaussian point clouds at extrapolated views are visualized at Fig.A2 in the revision.

This does not conflict with our claim. In the extrapolation setting, the geometry gaps are large rather than small refinements in precision, and large geometry errors do hurt rendering.

R2-W2: Controlled comparisons against LongLRM++ and MVP. Unified table.

(i) LongLRM++ FLOPs. We re-implemented its backbone exactly as specified in Sec. 4 of the paper. The parameter count matches the reported 110M. We will label the 700T as measured on our re-implementation. (ii) Decoder FLOPs. The reviewer is right that 56T excludes the decoder. The same holds for LongLRM++, whose 700T also counts reconstruction only. (iii) Recipe. Stages 1–4 of AsySplat-110 follow LongLRM++ exactly ([R1-W2]), so the stage-4 comparison isolates the backbone. With limited compute, we could not also follow MVP's protocol, which uses nearly twice as many training iterations as ours. (iv) MVP. MVP's FLOPs are comparable to ours. Against MVP, we claim fewer parameters and Gaussians, not lower compute. A unified table of inference overhead is given in [R1-W2/Q3].

R2-Q1: Numbers on the demo page.

Please refer to [R1-W2].

R2-Q2: Difference from iLRM.

Both methods avoid full attention over all high-resolution image tokens, but differ in what they decouple. (i) iLRM decouples the scene representation from the image tokens. A compact set of scene tokens iteratively aggregates information from the image tokens of all views and is refined across layers. (ii) AsySplat decouples the modeling of appearance and geometry based on their required granularity and learning difficulty. Inter-view interaction is restricted to coarse geometry tokens processed by a branch with larger capacity, fine appearance tokens are processed within each view by a lightweight branch.

R2-Q3: Qualitative comparisons with MVP.

Please refer to the Fig.A1 in the revision. We will update the video on the demo page.

AuthorsOpenReview ↗

Thank You and a Gentle Follow-up

Dear Reviewer B4tq,

Thank you again for your constructive comments. In response to your questions and suggestions, we have added experiments and clarifications to address your concerns about geometric accuracy, along with more comprehensive comparisons.

With less than two days remaining in the discussion period, we would like to kindly follow up to see whether you have any remaining concerns or would find further clarification helpful. We would be happy to provide additional explanations. Thank you sincerely for your time and thoughtful feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Thank You and a Gentle Follow-up Before the Discussion Deadline

Dear Reviewer B4tq,

We sincerely thank you again for your constructive comments. Following your suggestions, we have provided point-by-point responses to all of your questions and concerns, supported by additional experiments and detailed clarifications.

With less than 4 hours remaining in the discussion period, we would be grateful if you could review our rebuttal and share a brief follow-up comment. We would be happy to address any remaining questions or concerns.

Thank you again for your time and thoughtful suggestions.

Best regards,

The Authors

Reviewer_pf7xOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AsySplat studies efficient feed-forward 3D Gaussian reconstruction from posed, high-resolution, wide-coverage images. A higher-capacity geometry branch processes coarse tokens with cross-view interactions, while a lightweight appearance branch processes finer tokens within each view. Bilateral connections exchange information between the branches, whose features are fused for Gaussian prediction. Grid-based sparse attention further reduces computation. The paper evaluates a 98M base model against LongLRM and a 110M variant incorporating the LongLRM++ decoder and a revised training recipe. Its central argument is that allocating computation according to the different requirements of geometry and appearance improves the quality-efficiency trade-off.

Critical strengths

1.The base model provides a useful quality-cost trade-off across several datasets. Tables 2-3 report 98M versus 142M parameters and 57 versus 131 reconstruction TFLOPs, with comparable DL3DV-140 quality. Feed-forward PSNR improves over LongLRM by 0.65, 0.68, 0.55, and 1.59 dB on Tanks&Temples, MipNeRF-360, Deep Blending, and Replica, respectively; SSIM and LPIPS also improve across these datasets. Table 2 additionally evaluates quality under a fixed post-optimization time budget. This cross-dataset consistency supports the practical value of the proposed system.

2.The architecture has substantive ablation support. Table 5c compares single, symmetric, and asymmetric branches under two backbones. Tables A3a and A3b examine capacity allocation and reduced-size LongLRM variants, while A3d checks structural components at the target resolution. These controls provide evidence for the usefulness of the architecture and address plausible simpler alternatives. Their limited training scale restricts the strength of the broader conclusions, but they should receive substantial weight in assessing the contribution.

Critical weaknesses

1.The evidence for the central allocation strategy is not yet strong enough to establish the methodological contribution. The main novelty lies in how computation is allocated between geometry and appearance. Table 5c gives empirical support for this choice, but changes geometry-token granularity, computation, and potentially output Gaussian density together. Under Eq. (4), changing the patch sizes from (pc,pf)=(8,8) to (16,8) reduces the Gaussian count fourfold unless the prediction head is separately adjusted. The counts and capacities of these variants are not fully reported. Producing fewer Gaussians is a legitimate efficiency benefit; the unresolved issue is how much of the observed trade-off specifically supports the proposed allocation principles. The functional interpretation also needs care. Eq. (3) passes cross-view information into the appearance branch, and Eq. (4) combines both branches before attribute prediction. This is a coherent design, but it does not independently establish that appearance learning is intrinsically easier or can dispense with cross-view information. Table 5c does show that better measured geometry need not improve rendering in the tested setting. The broader explanation remains less well supported, particularly because this comparison uses the smaller 256x256 training regime. A3d verifies high-resolution component ablations, but does not repeat the central allocation/granularity comparison at the target operating point. There is a further uncertainty in the claimed controlled comparison. AsySplat's Eq. (5) specifies image loss and downsampled Depth Anything V2 supervision, whereas the original LongLRM paper includes opacity regularization and uses the original Depth Anything. These documented differences do not establish that the implemented comparison is unfair. They do require clarification before treating the full improvement as an architecture-only effect. If the evaluated models retain different supervision, a focused matched-supervision control is needed. Taken together, the current controls demonstrate a promising architecture, but leave too much uncertainty about the specific contribution of the allocation strategy.

2.The enhanced-model comparison does not close this evidence gap, and its practical efficiency is incompletely characterized. AsySplat-110 combines the borrowed decoder with changes to view curriculum, training duration, and depth supervision. Its competitive results establish the performance of this combined configuration. They do not show how much the proposed backbone contributes under that stronger recipe relative to a suitable alternative allocation. The feature-Gaussian prediction head and retained LongLRM++ components are also insufficiently specified. A controlled comparison using a shared decoder and training recipe would be more informative about the architectural contribution than the current base-to-enhanced comparison.The detailed timing and memory measurements concern the 98M model. For the 110M variant, reconstruction latency, per-target rendering cost, peak memory, and complete training cost are not reported. Relative to MVP, the reported reconstruction FLOPs are 56T versus 53T, although parameter and Gaussian counts are substantially lower. The reconstruction-FLOP comparison with LongLRM++ is explicitly labelled as an estimate and should not be treated as erroneous simply for that reason. Nevertheless, these measurements establish specific resource savings rather than a complete account of the enhanced system's practical efficiency. The supplied demo reports AsySplat-110 results of 26.13/0.834/0.184, while Tables 4 and A2 report 26.40/0.841/0.177. Different checkpoints or an outdated demo could explain this discrepancy. The authors should identify the authoritative configuration and evaluation results.

Questions for authors

1.For Table 5c, what are the parameter counts, output Gaussian counts, and prediction heads of the three variants? Is there a focused comparison that changes computation allocation while holding the relevant output budget and training recipe fixed, preferably at the target resolution? Please also state the depth reference and evaluation alignment used for the geometry metrics.

2.For Table 2, which LongLRM checkpoint and objective were used? Please clarify whether the depth estimator, depth-target resolution, and opacity regularization were matched. If the evaluated supervision differs, what evidence separates that effect from the architectural improvement?

3.Please provide AsySplat-110's reconstruction latency, per-target rendering latency/FPS, and peak memory alongside available baselines on the same hardware. Clarify the FLOP-counting convention and operation coverage, and distinguish its training cost from the estimate for the 98M model.

4 comments
AuthorsOpenReview ↗

Official Comment by Authors

We sincerely thank you for your constructive comments and suggestions. Below, we provide clarifications and additional experimental results, responding to the weaknesses (W) and the questions (Q) in your review.


R1-W1 / Q1: Focused comparison of Tab.5c. High-res ablations. Depth reference. Is appearance learning easier?

Focused comparison. We retrain the three variants of Tab.5c with matched parameter counts, output Gaussians and exactly the same training recipe. Two branches (8,8) applies two stride-2 convolutions to both branches before merging. The single branch is also processed with a stride-2 convolution before prediction.

256×256#ParamRecon. TFLOPsPred. head#GSPSNR↑δ<1.25↑
Single branch (p=8)101M14.7stride-2 conv -> (256/16)², 64 per token0.52M19.20.53
Two branches (8,8)100M13.4stride-2 conv on both branches -> (256/16)², 64 per token0.52M19.30.57
Asymmetric (16,8) (Ours)98M2.9stride-2 conv on appearance branch -> (256/16)², 64 per token0.52M19.30.52

The table shows that, with the parameters, output budget and recipe fixed, finer geometry tokens give more accurate geometry, but do not improve rendering.

High-res ablations. We finetune Asymmetric (16,8) and Two branches (8,8) from their 256² checkpoints to 540×960 under the same training protocol as Tab.A3d, and evaluate rendering results and depth on DL3DV-140. The results below also show that improving geometry within a limited range does not improve rendering on high-resolution input.

540×960Recon. TFLOPsPSNR↑δ<1.25↑
Two branches (8,8)48017.70.516
Asymmetric (16,8) (Ours)5717.70.462

Depth reference and alignment. In the main paper, pseudo-GT depth is obtained from MVSAnywhere prediction masked with MoGe, without per-image median alignment.

Is appearance learning easier? Please refer to [R3-Q2].

R1-W1 / Q2: Which LongLRM checkpoint and objective? Was supervision matched?

We use the officially released LongLRM checkpoint for evaluation, which is trained with the losses and weights listed below.

LongLRM (official)AsySplat-98M
L2 / perceptual weight1.0 / 1.01.0 / 0.5
opacity regularizer0.10
depth supervision0.01, Depth Anything v1, at full resolution0.01, Depth Anything v2, downsampled 2×

We agree that a focused matched-supervision control is needed. The following four models are re-trained with identical data, training schedule, and setups (256×256, 10k iters, batch24), and evaluated on DL3DV-140 (PSNR / SSIM / LPIPS):

LongLRM lossAsySplat loss
LongLRM (142M)18.78 / 0.501 / 0.51718.92 / 0.506 / 0.520
AsySplat (98M)19.10 / 0.517 / 0.49419.20 / 0.523 / 0.493

Under either loss, AsySplat outperforms LongLRM, while using 30% fewer parameters. The advantage over LongLRM is therefore architectural, not a product of supervision.

R1-W2: Controlled comparison using a shared decoder and training recipe. Demo error.

AsySplat-110 follows LongLRM++ in its decoder design (re-implemented following the paper as faithfully as possible), supervision, and training recipe for stages 1–4 (see the table below). The comparison at the end of stage 4 therefore provides a controlled evaluation of the contribution of the proposed efficient backbone.

StageAsySplat-110 (res. / iters / batch)LongLRM++ (res. / iters / batch)MVP (res. / iters / batch)
1256², 8 views / 60k / 256256², 8 views / 60k / 256270×480, 32 views / 100k / 256
2512², 8 views / 10k / 64512², 8 views / 10k / 64540×960, 32 views / 50k / 64
3540×960, 8 views / 10k / 64540×960, 8 views / 10k / 64540×960, mixed 16/32/64/128 views / 50k / (32~128) → 25.97(PSNR)
4540×960, 32 views / 10k / 64 → 26.27(PSNR)540×960, 32 views / 10k / 64 → 26.43(PSNR)—
5540×960, mixed 16/32/64/128 views / 20k / 64 → 26.40(PSNR)——
TFLOPs / Param56 / 110M~700 / 110M53 / 241M
Total iterations110k90k200k

Demo error. We apologize for the confusion. The numbers on the demo page were not updated when the paper was finalized. The paper's numbers are authoritative.

AuthorsOpenReview ↗

Official Comment by Authors

R1-W2 / Q3: Overhead.

All methods are evaluated on the same H200 GPU with the same gsplat version.

FLOP-counting convention. Forward-only FLOPs (2×MACs), counted with PyTorch's FlopCounterMode (matmul, conv and SDPA attention). The FLOPs of the fused Mamba2 scan are added analytically. All methods are counted with the same tool, and none of the counts include rasterization.

Method#Param#GSRecon. TFLOPsRecon. latencyRecon. peak memDecoder TFLOPs/viewRender ms/viewRender peak mem
LongLRM142M8M1310.54 s10.3 GB05.46.9 GB
iLRM186M4M2260.81 s7.0 GB05.37.3 GB
MVP241M16M530.18 s8.6 GB05.78.9 GB
AsySplat-98M98M4M560.36 s7.9 GB04.23.9 GB
AsySplat-110110M4M560.39 s7.6 GB0.5510.63.3 GB
LongLRM++110M4M~700——~0.55——

(i) Although AsySplat has FLOPs comparable to MVP, its latency is higher. This is because Mamba kernels are currently less hardware-optimized than attention. We adopt Mamba mainly for its lower training cost. (ii) The extra rendering cost of AsySplat-110 comes from the feature-Gaussian rendering inherited from LongLRM++, not from our backbone. AsySplat-98M, which shares the same backbone without this feature-Gaussian rendering, renders the fastest among all methods. (iii) Training AsySplat-110 takes 4.7k GPU hours, and AsySplat-98M takes 1.7k.

AuthorsOpenReview ↗

Thank You and a Gentle Follow-up

Dear Reviewer pf7x,

Thank you again for your constructive comments. In response to your questions and suggestions, we have added a series of experiments and clarifications. We hope these additions help address your concerns.

With less than two days remaining in the discussion period, we would like to kindly follow up to see whether you have any remaining concerns or would find further clarification helpful. We would be happy to provide additional explanations. Thank you sincerely for your time and thoughtful feedback.

Best regards,

The Authors

AuthorsOpenReview ↗

Thank You and a Gentle Follow-up Before the Discussion Deadline

Dear Reviewer pf7x,

We sincerely thank you again for your constructive comments. Following your suggestions, we have provided point-by-point responses to all of your questions and concerns, supported by additional experiments and detailed clarifications.

With less than 4 hours remaining in the discussion period, we would be grateful if you could review our rebuttal and share a brief follow-up comment. We would be happy to address any remaining questions or concerns.

Thank you again for your time and thoughtful suggestions.

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.