acceptodds
Under review as a conference paper at ICLR 2027

Grow Where Views Agree: Efficient Sparse-View 4D Reconstruction with GI-ACE

Abstract

With only four cameras, a dynamic Gaussian-splatting reconstruction sees each surface from few directions, yet densification still adds primitives wherever the pooled image-space gradient is large. We argue that this signal conflates two quantities: how strongly a primitive asks to change, and how many independent viewpoints make the request. A residual that only one camera sees can be removed by fitting that camera, so it is weak evidence that new scene structure is needed. We introduce GI-ACE (Geometry-Informed Asynchronous Camera Evidence), an admission rule for 4D Gaussian densification that uses the per-camera gradient statistics ordinary training already produces. A primitive grows if its demand survives removing its strongest camera, or, for a large primitive, if two cameras jointly request a split from sufficiently different directions, measured at the times each camera observed it. Both tests reuse the native threshold and add no learned component, point budget, or rendering pass. On the six-scene Neural3DV four-camera benchmark, GI-ACE exceeds, to our knowledge, all published results, improving mean PSNR from 22.29 to 23.04 dB over the strongest baseline, and in our reproductions it uses 2.9–3.7 fewer Gaussians while training 1.6–2.0 faster than the same backbone under native densification. On six Panoptic Sports sequences it improves mean PSNR by 0.45 dB with an order of magnitude fewer Gaussians and 5.4 faster training. Controlled studies confirm the design: the population stays stable under biased camera sampling, temporally staggered capture further improves the same rule, and the geometric route activates precisely where complementary views exist.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
4D Reconstruction, 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_ZVqFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GI-ACE, a camera-aware admission rule for densification in sparse-view 4D Gaussian reconstruction. It accumulates per-camera projected-gradient statistics during ordinary training. A robust core accepts a native growth proposal when its demand remains above the existing threshold after removing the strongest camera. For rejected large primitives, a geometric recovery branch admits splits supported by two sufficiently demanding and angularly separated views, using time-conditioned Gaussian positions. Experiments on Neural3DV and Panoptic Sports report substantially smaller Gaussian populations and shorter training times, alongside improved mean PSNR. Additional studies examine population-matched recovery, camera sampling imbalance, camera layout, and staggered observations.

Critical strengths

1.The paper makes a useful distinction between the magnitude of a densification request and the distribution of evidence across cameras. The strongest-camera-removal rule is simple and interpretable, and per-camera averaging avoids treating repeated observations from one camera as additional independent viewpoints. The method reuses existing backward-pass statistics without requiring another renderer, learned module, or explicit population budget.

  1. The reported efficiency gains are substantial across two datasets. On Neural3DV, the method reports higher mean PSNR than the published 4C4D result while reducing population and training time relative to the authors’ native-densification reproductions. On Panoptic Sports, the population and training-time reductions are particularly large. The sampling-imbalance experiments and explicit discussion of the two-observer boundary are useful analyses of how the rule behaves.
Critical weaknesses
  1. In Table 2, 4C4D’s image-quality scores come from the published paper, whereas its population and runtime come from the authors’ reproductions. The distinction is clearly disclosed, but the row does not describe a single baseline model. Since these reproductions already exist, their reconstruction scores should also be reported using the same evaluator as GI-ACE. Appendix D additionally notes that official and GI-ACE runs use different software environments, so the runtime ratios do not fully isolate the effect of the admission rule. A complete paired comparison is important for establishing the claimed quality–efficiency improvement.

  2. Table 4 retains the robust core and changes how additional recovery candidates are selected. This is a useful control for the geometric score, but it does not establish that the overall camera-aware policy is better than simpler ways of suppressing excessive growth. A higher native gradient threshold, a shorter densification schedule, or random admission over the full proposal pool could also reduce overfitting and population. Table 6 varies the threshold within GI-ACE rather than providing this native-densification comparison. The reported matched-count study therefore supports a narrower conclusion than the claim that the overall gains cannot be obtained by simply growing less.

  3. In Table 5, the canonical-ray variant achieves higher PSNR than the proposed time-conditioned variant under staggered observations: 24.737 versus 24.614 dB, with nearly identical populations and LPIPS. The paper acknowledges this result, which supports the method’s ability to operate with asynchronous evidence but does not demonstrate an advantage from evaluating rays at camera-specific observation times. The affine-trajectory derivation establishes how to compute a representative position; it does not establish that this position provides a better admission signal. This distinction matters because time-conditioned parallax is presented as a principal contribution.

Questions for authors
  1. Complete image-quality results for the native-densification reproductions in Table 2 would make the main comparison directly interpretable. Reporting quality, population, and runtime from the same runs, together with timing under a shared software environment, would clarify the contribution of GI-ACE itself.

  2. A native-densification quality–population curve obtained by varying its threshold or densification duration would help establish the benefit of camera-aware admission. A population-matched random-admission control over the full proposal pool would complement the recovery-only controls already provided.

  3. A broader comparison of canonical and time-conditioned rays, particularly on sequences with substantial motion, would clarify when the proposed temporal anchoring helps. The current staggered experiment supports asynchronous operation, but leaves the specific contribution of camera-dependent observation times unresolved.

2 comments

Response to Reviewer ZVqF

We thank the reviewer for the precise, actionable requests; we ran every control asked for. All numbers use the full held-out protocol. In short, no 4C4D operating point, at any threshold or schedule, reaches GI-ACE's quality at its population.

Z1. Complete paired comparison

The published rows keep the published 4C4D scores; the controlled comparison runs released 4C4D and GI-ACE on the same views and evaluator.

Scene4C4D: PSNR/LPIPS/GGI-ACE: PSNR/LPIPS/G
cook_spinach24.249 / 0.116 / 4.19M24.435+/-0.16 / 0.118 / 1.38M (3 seeds)
sear_steak23.66 / 0.118 / 1.74M24.520 / 0.117 / 0.57M
flame_steak23.72 / 0.114 / 2.46M24.685 / 0.113 / 0.80M
cut_roasted_beef23.22 / 0.124 / 3.63M24.101 / 0.123 / 0.76M
flame_salmon19.15 / 0.205 / 3.49M20.434 / 0.179 / 1.03M
Mean22.80 / 0.135 / 3.10M23.635 / 0.130 / 0.91M (3.4x fewer)

Our re-runs of the released code are close to the published 4C4D results; G counts are our reproductions. Uncapped 4C4D densification ran out of memory on 32 GB for 2 of 3 cook seeds and on salmon.

The controlled comparison gives better quality at 3.4x fewer Gaussians. GI-ACE exceeds 4C4D's quality (+0.84 dB mean, winning PSNR on all five scenes and LPIPS on four) with ~2x less training (5.0 h vs 2.2 h mean), and unrestricted 4C4D densification does not fit in 32 GB for 2 of 3 cook seeds: the removed capacity is the difference between trainable and untrainable on commodity hardware. Where a scene favors more capacity, the camera-evidence bar is a tunable tightness parameter: on cook, half-threshold admission reaches 24.928 dB at 3.13M, +0.68 dB above the 4C4D result (24.25 at 4.19M); on salmon it wins both axes outright (20.43/0.179 vs 19.15/0.205) and the loose bar improves LPIPS further to 0.155. The reverse, growing far less at fixed quality, 4C4D cannot do (Z2).

On the software stack: training time follows population. The 4C4D 2x-threshold run at 1.53M takes 8,782 s vs GI-ACE 8,861 s at 1.38M, so the advantage comes from admitting fewer Gaussians.

Z2. "Simply growing less?"

No. At no threshold or schedule does a 4C4D operating point reach GI-ACE's quality at its population. Below are the 4C4D threshold and schedule curves (also: stopping growth at 10k/5k gives 23.845/22.404; 1.0M cap 23.409; 8x threshold 21.563), plus the full-pool random control:

PolicyPSNRLPIPSFinal G
4C4D, default (full)24.2490.1164.19M
4C4D, 2x threshold24.0420.1391.53M
4C4D, hard cap at GI-ACE's 1.38M23.8490.1611.40M
4C4D, 4x threshold22.9210.1750.41M
Random admission, full pool (3 seeds)23.665+/-0.090.1501.68-1.87M
GI-ACE (3 seeds)24.435+/-0.160.1181.38M
GI-ACE, tighter (2x bar)23.9320.1570.39M
GI-ACE, looser (half bar)24.9280.1023.13M

Every 4C4D point under 1.9M is worse than GI-ACE by at least 0.39 dB, and even at 3.0x more Gaussians (4.19M) the full 4C4D run still trails it. At half threshold GI-ACE exceeds the 4C4D endpoint on both PSNR (+0.68) and LPIPS (0.102 vs 0.116) at 75% of its population: the GI-ACE frontier lies strictly above the 4C4D frontier.

Drawing the same number of proposals uniformly at random from the full proposal pool, without the robust core, loses 0.77 dB and 0.032 LPIPS with 22-35% more points (sear: -1.09 dB, 55% more). Under matched counts, the result depends on which proposals are admitted: camera-aware selection, not capacity.

Camera-imbalance stress test. Oversampling cam01 4x makes 4C4D quadruple its population to 4.79M while GI-ACE matches its quality at 4.2x fewer Gaussians (24.446 vs 24.482 dB). At 8x its population stays flat near 1.2M and the loss is only ~0.15 dB (24.29-24.32); dropping the normalization recovers the margin at +60% points (24.49 at 1.99M), a trade the loose bar dominates (24.93 at 3.13M).

Z3. Canonical versus time-conditioned rays

The two rules agree on 98.2% of admissions across the 294 events of an asynchronous run (7.5M evaluations): the ray convention barely changes which proposals are admitted, so the choice is second-order next to the admission rule itself (Z2). Temporal anchoring is what makes the score well-defined in the asynchronous regime, where each camera observes a different scene state and a canonical ray has no single instant to evaluate.

In the revision, time-conditioned parallax is reframed as a lightweight consistency formulation for asynchronous capture; the principal contribution is camera-aware admission, which Z2 shows drives the result.

In the revision we will separate the two tables and report the controls above.

Dear Reviewer ZVqF,

I hope this message finds you well. As the discussion period is nearing its end, we wanted to follow up and ensure our response has addressed your questions satisfactorily.

To briefly recap what we provided for your concerns: the requested controlled evidence is now delivered as dedicated tables: a quality–population frontier sweeping the baseline's own operating points, and a related-criteria comparison against Pixel-GS, AbsGS, and a fixed-budget MCMC baseline. On the mechanism question, we reframed time-conditioned parallax as a lightweight consistency formulation for asynchronous capture rather than a principal contribution, and the logged decisions show the two ray conventions agree on 98.2% of admissions. The camera-aware admission rule remains the paper's central claim.

If there are any additional points or feedback you'd like us to consider, please let us know; we are happy to provide further details or controls. Your scrutiny on experimental protocol genuinely improved the rigor of the revision.

Thank you for your time and effort in reviewing our work.

Best regards, Authors

Reviewer_o3UROpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GI-ACE, a method for making sparse-view 4D Gaussian reconstruction more efficient by improving how new Gaussians are added during training. The main idea is that a large gradient from one camera doesn't necessarily mean the scene needs more Gaussians, so the authors try to use agreement across different camera views to make better densification decisions. Specifically, the method first checks whether a growth request is still strong after removing the camera with the largest gradient contribution. It then uses geometric information, including the viewing angles and observation times of different cameras, to recover potentially useful splits. The method doesn't require additional training modules or extra rendering passes. Experiments on Neural3DV and Panoptic Sports show improvements in PSNR while using significantly fewer Gaussians and reducing training time. Overall, the paper explores an interesting way to make Gaussian densification more selective by considering multi-view evidence rather than just gradient magnitude.

Critical strengths

I think the paper starts from a pretty interesting observation: a large gradient doesn't necessarily mean we need more Gaussians, especially when most of that signal comes from just one camera. Using agreement across different views to guide densification seems like a reasonable idea, and I like that the authors turn this into a fairly simple method without introducing extra networks or rendering passes. The experimental results are also quite encouraging, especially the large reduction in Gaussian count and training time while still improving PSNR. I also appreciate that the authors go beyond standard benchmark comparisons and include matched-count recovery, camera sampling imbalance, and asynchronous observation experiments. These experiments help explain why the method works rather than just showing better numbers. Overall, I think the paper has a clear motivation, a practical design, and some meaningful efficiency improvements.

Critical weaknesses

My main concern is the geometric reasoning behind the time-conditioned parallax. Since different cameras may observe a moving Gaussian at different timestamps, the two viewing rays don't necessarily point to the same 3D location. So I'm not fully convinced that a larger angle always means stronger geometric evidence, as it could also come from object motion. Interestingly, the canonical-ray variant in Table 5 actually achieves slightly better PSNR, which makes me question how necessary the time-conditioned design really is. I also think the contribution of the geometric recovery module is not entirely clear, since the improvement over the robust core alone seems relatively small, and most experiments only report single runs. Another concern is the baseline comparison on Neural3DV, where the quality results come from the published paper while runtime and Gaussian counts come from the authors' own reproductions. It would be more convincing to compare everything under exactly the same setup. Also, although the PSNR improvements are promising, LPIPS actually gets worse across all six Panoptic Sports sequences, so the quality improvement is not as consistent as it first appears. I think the overall idea is interesting, but the paper still needs stronger evidence to support some of its key design choices.

Questions for authors

I have a few questions about the method and experiments. First, since different cameras may observe a moving Gaussian at different times, how can we be sure that the angle between their viewing rays reflects actual geometric constraints rather than just object motion? Also, I noticed that the canonical-ray variant in Table 5 achieves slightly better PSNR than the proposed time-conditioned version. Could the authors explain what the time-conditioned design actually contributes here? I'm also curious about how much improvement comes specifically from geometric recovery, since the gain over the robust core alone seems rather small. Would this improvement remain consistent across different scenes and random seeds? Finally, could the authors report the reconstruction quality of their reproduced 4C4D baseline under exactly the same evaluation setup? I think this would make the quality and efficiency comparisons more convincing.

2 comments

Response to Reviewer o3UR

We thank the reviewer for the positive assessment and the sharp questions. In short, GI-ACE matches the strongest published baseline (4C4D) at 3.6x fewer Gaussians and ~2x less time; a looser bar passes its best point.

Q1. Can the ray angle reflect motion rather than geometry?

Correct: rays toward a moving primitive at different times need not meet at one 3D point, so a large angle alone cannot certify correspondence. The angular factor is a directional-complementarity heuristic on the estimated trajectory, bounded by three properties:

  1. Recovery only, never invention: it only re-admits splits the base 4C4D densifier itself requested and the robust core rejected, each still demanded by two cameras above the densification threshold.
  2. Bounded angle effect: the sine factor only reweights candidates inside this pool, degrading at worst to a demand-only pair test.
  3. Consistency on the trajectory: rays are evaluated at each camera's gradient-weighted observation time on the trajectory.

What matters is where growth is admitted: same-count random admission from the full pool drops 0.77 dB on cook and 1.09 dB on sear, with more Gaussians.

Q2. What does the time-conditioned design contribute?

On standard capture the contribution is small: across the 294 events of an asynchronous run (7.5M evaluations), the two rules agree on 98.2% of admissions, and a same-seed canonical control gives 24.715 vs 24.375 time-conditioned, a second-order detail. Temporal anchoring matters only under asynchronous capture, where canonical evaluation is ill-defined; there it lands on either side of canonical (24.614 / 24.022 vs 24.737). It is a lightweight consistency formulation for asynchronous capture.

Q3. How much does recovery add? Is it stable across seeds?

Recovery adds +0.24 dB on the paired seed, consistently across the three seeds. On cook (seeds 42/7/123, same-seed core-only control): core-only 24.139 / 1.34M; GI-ACE (core + recovery) 24.435 +/- 0.16 / 0.118 / 1.38M; random recovery at matched count 24.454 +/- 0.07 / 0.119 / 1.30-1.50M.

  1. Recovery helps consistently: 24.375 vs 24.139 on the paired seed (+0.24 dB at +0.04M); all three seeds (24.31-24.62) sit above core-only.
  2. Its main role is setting how many splits to recover; given that count, random selection performs similarly.
  3. The camera-aware core is essential: removing it costs 0.77 dB even with more Gaussians.

Q4. Reconstructed baseline under exactly the same setup

The paired comparison gives matched quality at 3.6x fewer Gaussians. We ran released 4C4D under the same configuration, views, and evaluator on the 5 N3V scenes: 4C4D 23.601 / 0.123 / 3.29M / ~4.9 h vs GI-ACE 23.635 / 0.130 / 0.91M / ~2.2 h. Per scene, GI-ACE ties or wins PSNR on cook, sear, steak, and cut (+0.48 dB), losing only on salmon at 3.4x less capacity, where it still beats official-config 4C4D outright (20.43 vs 19.50, LPIPS 0.179 vs 0.199) and on cook likewise (24.44 vs 24.25 at 3x fewer). Uncapped 4C4D ran out of memory (32 GB) for 2 of 3 cook seeds: the over-growth is infeasible outright.

The trade-off is a tunable knob: the GI-ACE frontier sits above the 4C4D frontier. Loosening on cook (half threshold) gives 24.928 dB at 3.13M, above the 4C4D endpoint's 24.365 at 5.10M, with the best LPIPS 0.102; tightening (2x threshold) gives 23.932 at 0.39M, still above 4C4D's matched point (22.921 at 0.41M). On salmon the same knob recovers perceptual quality past 4C4D (0.155 vs 0.169).

Training time tracks population: at matched population 4C4D (1.53M) and GI-ACE (1.38M) agree within 1% (8.8k vs 8.9k s).

Under imbalanced capture, oversampling cam01 4x makes 4C4D quadruple its population to 4.79M while GI-ACE matches its quality at 4.2x fewer Gaussians (24.446 vs 24.482 dB). At 8x the population stays flat near 1.2M and the loss is only ~0.15 dB (24.29-24.32); dropping the normalization recovers the margin at +60% points (24.49 at 1.99M), a trade the loose bar dominates (24.93 at 3.13M).

Q5. LPIPS on Panoptic Sports

At equal capacity GI-ACE's LPIPS beats 4C4D's; the Panoptic gap is a property of that dataset. On cook at matched population GI-ACE's LPIPS is clearly better (0.118 at 1.38M vs 0.139 at 1.53M vs 0.161 capped at the same 1.38M), and the looser bar reaches 0.102, better than even the 5.10M 4C4D run (0.111) - the best score of any row. Rankings hold under the standard [-1,1] convention.

On Panoptic vs released-code 4C4D (4M cap reached on all six): 4C4D 18.14 / 0.160 DSSIM / 0.334 LPIPS / 4.00M / 3.8 h vs GI-ACE 18.09 / 0.143 DSSIM (6/6 better) / 0.367 / 0.29M / 0.54 h (7x faster). LPIPS is worse on all six, but the gap is a property of the scene: on football (largest gap) growing the population 4.5x (0.31M to 1.38M, 2.9x below 4C4D) left LPIPS unchanged (0.373) with PSNR above 4C4D (18.17 vs 17.96; official: 16.47). DSSIM better on all six with level PSNR: fine texture preserved at 14x less capacity.

Dear Reviewer o3UR,

I hope this message finds you well. As the discussion period is nearing its end, we wanted to follow up and ensure our response has addressed your questions satisfactorily.

To briefly recap what we provided for your central question, whether our gain comes from simply growing less: a new quality–population frontier table sweeps the baseline's own threshold, cap, and growth-schedule operating points, and every point below 1.9M loses at least 0.60 dB to GI-ACE, while GI-ACE's loosest point exceeds the densest baseline run on both PSNR and LPIPS at 75% of its population. Random admission from the full native pool reaches only 23.67 dB while spending 45–61% more Gaussians, and the robust core alone reaches 24.14 dB, so both stages of the rule contribute. The matched-count recovery controls you asked about isolate selection from capacity at equal population.

If there are any additional points or feedback you'd like us to consider, please let us know; we are happy to run further analyses. Your questions pushed us to make the evidence much stronger, and the revision is better for it.

Thank you for your time and effort in reviewing our work.

Best regards, Authors

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

This paper addresses the problem of insufficient Gaussian densification in sparse view dynamic reconstruction. They introduce a geometry-informed densification strategy with two component: a strongest-camera-removal test and a geometric route that recover a split supported by two cameras using demand-weighted parallax. The experimental results demonstrate improvement in psnr with better efficiency.

Critical strengths
  • The paper is well motivated by the limitation of conventional gradient-based densification for sparse-view supervision.
  • The proposed strategy is simple and intuitive, achieving improved rendering quality while reducing Gaussian counts and training time.
Critical weaknesses
  • The geometric admission criterion is not fully justified. The proposed angular separation does not necessarily indicate consistent geometric evidence, particularly when the primitive moves between observations. The experiments do not fully verify that this criterion selects geometrically more accurate primitives. Moreover, Table 5 shows that using canonical-center rays achieves higher PSNR than the proposed time-conditioned rays in the asynchronous setting, raising questions about the necessity of the temporal formulation.
  • Lack of comparisons with related works. The proposed work mainly revises the Gaussian densification strategy, but lacks comparison with similar works in this area, like AbsGS [1], Pixel-GS [2], 3DGS-MCMC [3], and Revising densification [4]. The paper should compare with these works using the same backbone.
  • The improvement in reconstruction quality is not consistent. As shown in Table 3, while the PSNR and SSIM are improved, the LPIPS is not better. This may suggest that the densification removes high-frequency details and produces blurry rendering. This should be further explained and analysed.

[1] Ye, Zongxin, et al. "Absgs: Recovering fine details in 3d gaussian splatting." Proceedings of the 32nd ACM international conference on multimedia. 2024.

[2] Zhang, Zheng, et al. "Pixel-gs: Density control with pixel-aware gradient for 3d gaussian splatting." European conference on computer vision. Cham: Springer Nature Switzerland, 2024.

[3] Kheradmand, Shakiba, et al. "3d gaussian splatting as markov chain monte carlo." Advances in Neural Information Processing Systems 37 (2024): 80965-80986.

[4] Rota Bulò, Samuel, Lorenzo Porzi, and Peter Kontschieder. "Revising densification in gaussian splatting." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2024.

Questions for authors

See weaknesses.

3 comments

Response to Reviewer uCdE

We thank the reviewer. The new experiments below address each point; all numbers use the full held-out protocol. In short, GI-ACE matches or beats the strongest published baseline (4C4D) at a fraction of the Gaussians, and a looser admission bar moves past 4C4D's best point.

W1. Is the angular factor a sound criterion, and does temporal anchoring drive the gains?

The angular factor is a bounded heuristic that decides how many splits to recover; the gains come from camera-aware selection. It only re-admits large splits that the base 4C4D densifier itself requested and the camera-aware core rejected, each still demanded by two cameras above the densification threshold. In paired-seed controls on cook (seeds 42/7/123), random selection at equal count is on par (24.454 +/- 0.07 vs 24.435 +/- 0.16 dB). What matters is camera-aware selection of where to grow:

cook_spinachPSNRLPIPSGaussians
GI-ACE (3 seeds)24.435 +/- 0.160.1181.38M
Random admission, full pool, same count, no core (3 seeds)23.665 +/- 0.090.1501.68-1.87M
4C4D densification, best point under 1.9M24.0420.1391.53M

Same-count random admission loses 0.77 dB with more Gaussians, and 1.09 dB on sear_steak (23.430 at 0.88M vs 24.520 at 0.57M). Camera-aware evidence drives the gain.

We logged which splits each rule would admit at the 294 densification events of an asynchronous run (7.5M evaluations, same model state): the rules agree on 98.2%. A same-seed canonical control on cook gives 24.715 dB vs 24.375 for time-conditioned rays, a second-order detail. Temporal anchoring matters only under asynchronous capture, where canonical evaluation is ill-defined; there it lands on either side of canonical (24.614 / 24.022 vs 24.737). We treat it as a lightweight consistency formulation for asynchronous capture.

W2. How does GI-ACE compare with related densification methods?

GI-ACE exceeds every related criterion on the same backbone. We implemented the Pixel-GS, AbsGS, and MCMC-style criteria in the released 4C4D code, same representation, loss, views, and evaluator.

cook_spinachPSNRLPIPSGaussians
4C4D densification24.3650.1115.10M +
Pixel-GS22.7490.1750.46M
4C4D, hard cap at 1.38M23.8490.1611.40M
MCMC-style fixed budget @1.38M +23.5710.1381.39M
AbsGS (original two-criterion design) *23.7420.1560.89M
GI-ACE (3 seeds)24.435 +/- 0.160.1181.38M
GI-ACE, looser bar24.9280.1023.13M
  • Uncapped 4C4D ran out of memory on 32 GB for seeds 42 and 7; seed 123 shown. Under its official configuration 4C4D is weaker still on cook (24.25 / 0.116 / 4.19M); the row above is the stronger shared-configuration variant. + Dead-point relocation at a fixed population. * Per-view absolute-gradient accumulation (per-pixel needs a patched rasterizer); standard gradients gate cloning, homodirectional ones gate splitting.
  • Pixel-GS loses 1.7 dB.
  • AbsGS reaches 23.742 dB at 0.89M: 0.69 dB below GI-ACE, worse LPIPS, below even 4C4D capped at 1.53M (24.042). AbsGS changes what gradient statistic is measured; GI-ACE changes how evidence is pooled across cameras. The two are complementary.
  • Budgeted methods stay below GI-ACE at equal counts: 4C4D capped at the same 1.38M reaches 23.849 and MCMC relocation 23.571, 0.59 and 0.86 dB below. A fixed budget cannot say where to grow; camera evidence can.
  • At a looser bar GI-ACE reaches 24.928 / 0.102 at 3.13M, above AbsGS and the 5.10M 4C4D run at 39% fewer Gaussians.
  • With one camera oversampled 4x, 4C4D inflates to 4.79M while GI-ACE matches its quality at 4.2x fewer.

W3. Do the lower LPIPS scores indicate blurring?

(a) No: at matched population GI-ACE's LPIPS is already better than 4C4D's. On cook, GI-ACE reaches 0.118 at 1.38M vs 4C4D's 0.139 at 1.53M and 0.161 capped at the same 1.38M; the loose end reaches 0.102, better than the 5.10M 4C4D run (0.111).

(b) Structure improves as well. On Panoptic Sports vs released-code 4C4D (4M cap reached on all six): DSSIM better on all six (0.143 vs 0.160) with level PSNR (18.09 vs 18.14; +0.45 mean vs 4C4D's official settings), at 14x fewer Gaussians and 7x less training. On football, growing the population 4.5x (0.31M to 1.38M) left LPIPS unchanged (0.373) while PSNR stayed above 4C4D (18.17 vs 17.96; official: 16.47).

(c) The trade-off is a knob. On cook the bar spans 23.932 dB at 0.39M to 24.928 dB at 3.13M, above the 4C4D endpoint. On salmon it recovers perceptual quality past 4C4D (LPIPS 0.155 vs 0.169), and against 4C4D's official settings wins outright (20.43 vs 19.50 dB). Growing far less at fixed quality is what 4C4D cannot do at any threshold.

(d) Rankings hold under the standard [-1,1] convention (0.182-0.185 vs 0.208).

We hope the reviewer will reconsider in light of these controls.

Dear Reviewer uCdE,

I hope this message finds you well. As the discussion period is nearing its end, we wanted to follow up and ensure our response has addressed your questions satisfactorily.

To briefly recap what we provided for your concerns: the requested related-criteria comparison is now in the paper. On the same backbone and evaluator, Pixel-GS, AbsGS, and a fixed-budget MCMC baseline each land 0.80–1.90 dB below GI-ACE even at up to 21% larger populations. On the software-stack question, we showed wall-clock time is governed by population (8,782 s vs 8,835 s at matched populations), so the reported speedups are practical end-to-end comparisons. The robustness controls on sampling imbalance and staggered capture are also included.

If there are any additional points or feedback you'd like us to consider, please let us know; we are happy to run further analyses. Your careful reading has materially improved the paper.

Thank you for your time and effort in reviewing our work.

Best regards, Authors

Thanks for the response and the additional experiments. While the rebuttal strengthens the paper, several concerns still remain. First, I remain concerned about the necessity and effectiveness of the proposed geometric recovery criterion. The matched-count experiments show that random selection can achieve comparable performance (24.454 vs. 24.435 dB), weakening the claim that angular-aware candidate selection is essential. In addition, the consistent LPIPS degradation on Panoptic Sports remains insufficiently explained. Overall, although the rebuttal strengthens the empirical results, it does not fully resolve the concerns regarding the methodological contribution and the necessity of the proposed geometric design. So I will remain my original rating.

Comments on the paper

1 comment

Dear Area Chair,

We thank you and the three reviewers for the careful reading. Every substantive question raised is answered in the revision with new controlled experiments rather than prose adjustments; all changes are marked in blue and the revision stays within the ten-page limit. As a reminder, GI-ACE improves mean PSNR from 22.29 to 23.04 dB over the strongest published baseline on Neural3DV, and +0.45 dB on Panoptic Sports at 11x fewer Gaussians and 5.4x faster training. We summarize how each concern maps to a concrete change.

Evidence integrity (ZVqF). The published per-scene comparison is kept with provenance stated per column in the caption (quality transcribed from Zhou et al.; populations and times, which the published work never reported, are our reproductions on the same GPU). The reviewer's underlying request, single-run controlled evidence, is now met directly by two new controlled tables: a quality-population frontier sweeping the baseline's own threshold, cap, and growth-schedule operating points, and a related-criteria comparison. Uncapped 4C4D densification also fails to fit in 32 GB on two of three Cook Spinach seeds and on Flame Salmon, so the capacity discipline GI-ACE provides is a practical necessity for the backbone.

"Simply growing less" (o3UR Q1, ZVqF Z2). The new frontier table shows every 4C4D operating point below 1.9M losing at least 0.60 dB to GI-ACE, and GI-ACE's loosest point exceeds the densest 4C4D run on both PSNR and LPIPS at 75% of its population. Removing the camera-aware stage entirely is costlier still: the robust core alone reaches 24.14 dB, and random admission from the full native pool reaches only 23.67 dB while spending 45-61% more Gaussians (-0.98 dB). The matched-count recovery controls further isolate selection from capacity (24.65 vs 24.36-24.63 random at equal population). Both stages of the rule contribute, and the margin exceeds the measured seed-level spread several times over.

Related criteria (uCdE W2). Pixel-GS, AbsGS, and a fixed-budget MCMC baseline, the criteria suggested by the reviewer, are now compared on the same backbone and evaluator and cited in the paper: each lands 0.80-1.90 dB below GI-ACE even at up to 21% larger populations, and none improves its LPIPS. Camera-agnostic densification criteria do not recover the gain; it is specific to camera-aware admission.

Mechanism framing (ZVqF Z3). Time-conditioned parallax is demoted from a principal contribution to a lightweight consistency formulation for asynchronous capture; logged decisions show the two ray conventions agree on 98.2% of admissions, so the convention is a second-order detail relative to the admission rule itself.

Robustness (uCdE W3, ZVqF W5). Under 4x camera oversampling, released 4C4D inflates to 4.79M Gaussians while GI-ACE matches its quality at 4.2x fewer; the Panoptic LPIPS gap is shown to be capacity-invariant on two sequences; staggered capture improves GI-ACE to 24.61 dB while degrading the baseline; and wall-clock time is population-governed (8,782 s vs 8,835 s at matched populations), so the reported speedups are practical end-to-end comparisons.

All numbers above are consistent with the per-reviewer responses and appear in the revision. We hope the revised manuscript and the new controls answer the reviews to your satisfaction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.