acceptodds
Withdrawn from ICLR 2027

FormFlow-GS: A General and Efficient Method for Online Multimodal Gaussian Mapping

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Multimodal online Gaussian mapping still suffers from critical limitations in geometric formation and refinement efficiency. Although depth or LiDAR provides metric surface geometry, this information is often underused when new Gaussians are formed, while downstream refinement incurs substantial execution overhead. Addressing both stages is therefore essential for high-quality and efficient online mapping. We propose FormFlow-GS, a general Gaussian mapping method built around two components. GeoBirth uses local metric geometry and neighborhood reliability to form surface-aligned Gaussians with adaptive anisotropic support, providing a better-conditioned geometric starting point for refinement. Coalesced Gaussian Optimization (CGO) coalesces heterogeneous attribute updates over a shared visible-Gaussian workset, reducing redundant execution during refinement. The design is broadly applicable and can be readily integrated into Gaussian mapping backbones spanning monocular, RGB-D, RGB-LiDAR, and LiDAR-RGB-IMU settings. Across these backbones and multimodal benchmarks, it consistently improves rendering quality and online mapping efficiency, achieving a strong quality-efficiency balance and outperforming most evaluated state-of-the-art baselines overall.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Online Gaussian Mapping, Multimodal Mapping, LiDAR-Visual Mapping, Gaussian Formation, Efficient Optimization

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_eC7eOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes FormFlow-GS, an online multimodal 3D Gaussian Splatting (3DGS) framework designed to address geometric initialization errors and downstream optimization latency. The method introduces two main components: (1) GeoBirth, which utilizes local metric geometry from LiDAR or depth along with neighborhood reliability to initialize surface-aligned anisotropic 3D Gaussians; and (2) Coalesced Gaussian Optimization (CGO), a CUDA kernel fusion strategy that consolidates heterogeneous attribute updates (position, rotation, scale, opacity, and spherical harmonics) into a unified kernel pass across a shared visible-Gaussian workset. The framework is evaluated across monocular, RGB-D, and LiDAR-inertial SLAM/mapping backbones (including Gaussian-LIC2, GS-LIVM, PVG, SGAD-SLAM, VarSplat, and MonoGS) on several multimodal benchmarks.

Critical strengths

S1: Relevant system-level optimization focus for real-time Gaussian mapping. Modern online 3DGS systems frequently suffer from GPU thread divergence and excessive kernel launch scheduling overhead when updating decoupled Gaussian parameters sequentially across separate CUDA kernels. Restructuring the multi-attribute gradient accumulation into a coalesced execution path over active visible Gaussians targets a practical engineering bottleneck relevant to compute-constrained edge robotics and embedded mapping devices.

S2: Broad cross-backbone evaluation scope across sensor configurations. The authors make a noticeable effort to interface their approach with diverse sensor suites and baseline architectures, spanning monocular (MonoGS), RGB-D (SGAD-SLAM, VarSplat), and LiDAR-visual-inertial setups (Gaussian-LIC2, GS-LIVM, PVG), attempting to substantiate cross-modal transferability rather than restricting evaluation to a single sensor setup.

Critical weaknesses

W1: Misleading efficiency claims and neglect of Amdahl's Law in end-to-end mapping. Throughout the introduction and abstract, the paper heavily emphasizes major speedups and reduced optimization overhead, foregrounding micro-benchmark metrics such as kernel launches dropping from 63 to 1 (or 40 to 1) and isolated optimizer GPU execution time decreasing by 70% to 92%. However, an examination of the actual system-level wall-clock mapping time in Table 1 reveals that these micro-optimizations yield negligible end-to-end benefits across nearly all evaluated pipelines: VarSplat on TUM decreases by only 0.62% (from 206.72s to 205.44s), SGAD-SLAM on TUM decreases by only 1.99% (from 11.53s to 11.30s), and GS-LIVM decreases by only 2.30% (from 12.61s to 12.32s). In online SLAM and dense visual mapping, optimizer kernel scheduling occupies only a minor slice of the overall compute budget compared to pose tracking, spatial point cloud projection, voxel downsampling, visibility sorting, and differential rasterization. Highlighting isolated kernel fusion gains as a breakthrough in online mapping efficiency while total system runtime remains virtually unchanged constitutes an ungrounded claim.

W2: Statistically insignificant PSNR improvements accompanied by systematic degradation in perceptual quality (LPIPS metric inversion). Across standard benchmarks in Table 1, the observed rendering gains in PSNR are minimal—often within the margins of stochastic variance (+0.103 dB on VarSplat, +0.105 dB on SGAD-SLAM, and +0.154 dB on PVG). More critically, the perceptual metric LPIPS exhibits consistent negative optimization. In Table 1 on Gaussian-LIC2, applying FormFlow-GS degrades LPIPS from 0.2390 to 0.2501 (+0.0111 increase, where lower indicates better perceptual fidelity). This contradiction is reinforced by the 2x2 controlled decomposition in Table 2: the unmodified baseline E00 achieves the best LPIPS score of 0.2390, whereas adding GeoBirth (E10) degrades it to 0.2550, and the complete pipeline (E11) remains degraded at 0.2501. The text repeatedly claims enhanced boundary sharpness and finer visual textures, yet fails to report, acknowledge, or reconcile why the primary perceptual quality metric consistently deteriorates across configurations.

W3: Benchmarking against catastrophic baseline failures without technical justification. In Table 1, the reported baseline numbers for GS-LIVM (PSNR 6.175 dB, SSIM 0.3735) and MonoGS (PSNR 7.980 dB, SSIM 0.0517) represent completely collapsed, non-converged map states rather than functional, operational baselines (where typical novel-view rendering on these benchmarks exceeds 18 to 25 dB). Presenting an incremental gain on top of a failed system (such as moving from 7.980 dB to 9.610 dB, which remains an unusable reconstruction) does not validate the proposed method's capability. The authors provide no error analysis explaining why these baselines experienced catastrophic tracking or mapping loss, leaving it ambiguous whether the baselines were misconfigured or if divergent trajectories were retained simply to display numerical deltas.

W4: Conceptual fragmentation and absence of mathematical synergy. The paper attempts to frame GeoBirth and CGO as a mutually reinforcing, full-lifecycle mapping methodology where formation conditions refinement. In reality, GeoBirth is an intuitive geometry-based initialization that aligns Gaussian ellipsoids with local surface normals and covariance (a well-established strategy in methods like PGSR, 2DGS, and GeoGS), while CGO is an orthogonal, low-level CUDA operator fusion. There is no mathematical bridge or algorithmic dependency connecting the two components: CGO does not utilize geometric priors established by GeoBirth, nor does GeoBirth structure the coalesced memory layout of CGO. Packaging two disparate engineering steps under an abstract "formation-refinement" narrative artificially inflates the conceptual depth of the work.

W5: Unexplained optimization dynamics and numerical inconsistency in ablation studies. In the 2x2 ablation in Table 2, CGO is introduced as an execution-level kernel fusion that aggregates heterogeneous attribute updates. If CGO performs mathematically equivalent gradient updates across attributes, transitioning from E00 (Native+Native, 20.692 dB) to E01 (Native+CGO) should yield identical reconstruction fidelity. However, E01 drops to 20.675 dB, while combining CGO with GeoBirth in E11 produces an unexplained non-linear leap to 21.041 dB (+0.194 dB over E10). The paper contains no mathematical convergence analysis, discussion of floating-point reduction order, atomic write race conditions, or gradient clipping differences to explain why an operator fusion alters optimization trajectories, raising doubts about experimental reproducibility.

Questions for authors

Q1: How do the authors reconcile the paper's claims of sharper visual fidelity with the consistent degradation of LPIPS across Table 1 (Gaussian-LIC2) and Table 2 (where baseline E00 achieves 0.2390 while E11 worsens to 0.2501)? What visual artifacts or spatial smoothing are introduced by GeoBirth or CGO that cause perceptual fidelity to decline despite minor PSNR shifts?

Q2: Given that end-to-end mapping time in Table 1 decreases by less than 1% on VarSplat (0.62%), under 2% on SGAD-SLAM (1.99%), and under 2.5% on GS-LIVM (2.30%), what is the practical runtime benefit of CGO when integrated into full robotic pipelines where tracking, projection, and rendering dominate the compute budget?

Q3: What specific factors led to the baseline collapse of GS-LIVM (PSNR 6.175 dB) and MonoGS (PSNR 7.980 dB) in Table 1? Did the tracking frontend diverge, or were default parameters modified? Why are comparisons against unusable reconstructions considered valid benchmarks for cross-backbone transferability?

Q4: Is CGO mathematically equivalent to the native multi-kernel attribute update? If so, why does E01 deviate in PSNR and SSIM from E00 in Table 2? Does CGO introduce non-deterministic floating-point accumulation or thread contention during online optimization?

Q5: Why does the paper omit standard 3D surface reconstruction metrics (e.g., Chamfer Distance, F-score, or point-to-plane RMSE) when evaluating GeoBirth, relying solely on 2D novel-view rendering metrics to claim improved 3D metric surface alignment?

5 comments
AuthorsOpenReview ↗

Response to Reviewer eC7e (Q1): LPIPS Degradation and Reliability-Guided Shaping

In the 12-sequence evaluation on Gaussian-LIC², FormFlow-GS improves PSNR from 20.692 to 21.041 dB and SSIM from 0.7160 to 0.7216. To address the lack of a corresponding LPIPS improvement, we conducted controlled experiments that reveal the relationship between spatial support strength and perceptual detail. Across three test scenes, Reliability reduces mean LPIPS by approximately 12.53% relative to Full-strength, validating the effectiveness of adaptive shaping in mitigating perceptual loss. We thank the reviewer for pointing out this issue and analyze it through supplementary experiments.

1. Causes of perceptual degradation and controlled verification

GeoBirth changes the image-projection support of Gaussians by adjusting their anisotropic scales. Stronger spatial support can improve pixel fitting within a region, but may also increase projection overlap and color blending, smoothing high-frequency textures such as text strokes, thin lines, and local edges.

To verify this mechanism, we fixed 464,435 Gaussians, birth candidates, camera poses, and optimization conditions, varying only spatial support strength. The results show that PSNR increases from 21.2736 to 21.8861 dB (+0.6125 dB), whereas standard LPIPS increases from 0.33555 to 0.37535 (+0.03980), confirming that excessive spatial shaping may cause a tradeoff between pixel accuracy and perceptual quality. Further same-view comparisons are presented below.

Table 1: Rendering quality comparison for the full view and local text region

Evaluation regionMethodPSNR (dB) ↑SSIM ↑LPIPS ↓
Full imageNative E0020.2840.81270.4268
Full imageFull-strength20.4280.81610.4395
Full imageOurs E1120.4610.81500.4204
Local text region (5×)Native E0017.3890.73280.2405
Local text region (5×)Full-strength18.0150.69490.2646
Local text region (5×)Ours E1118.1570.74550.1930

Although Full-strength improves PSNR, it also degrades LPIPS. In contrast, Ours E11 achieves both higher PSNR and lower LPIPS in the full image and obtains the best results on all three metrics in the local text region. This indicates that Reliability-guided shaping can effectively preserve local perceptual detail in this test view⁠.

2. The role of Reliability in modulating perceptual quality

GeoBirth adaptively adjusts the anisotropic support of Gaussians according to local geometric reliability, strengthening surface alignment in geometrically trustworthy regions and reducing shaping strength where reliability is insufficient, thereby mitigating excessive smoothing caused by fixed full-strength shaping.

Table 2: Reliability ablation and changes in standard LPIPS for the full method

SceneFull-strength ↓Reliability ↓LPIPS change ↓
Bright Screen Wall0.375420.34896−0.02646
R3LIVE degenerate 010.302090.21097−0.09111
M2DGR room030.445660.42248−0.02318

Note: Negative values indicate an LPIPS improvement in the corresponding comparison.

Reliability outperforms Full-strength in all three scenes, with an average LPIPS improvement of approximately 12.53%, further validating the positive role of adaptively controlling spatial shaping strength in mitigating the loss of perceptual detail.

3. Does CGO introduce perceptual degradation?

We performed bitwise-consistency verification of 10,692 parameter and momentum updates, confirming that CGO and the tested native visibility-gated Adam produce identical updates under identical inputs. Further controlled experiments localized the first historical numerical difference to the Backward gradient-accumulation stage, before the optimizer. With the Backward configuration fixed, switching only CGO did not show stable changes in PSNR or LPIPS. Therefore, the available evidence does not support attributing perceptual degradation to CGO’s parameter-update rule.

In summary, GeoBirth improves Gaussian surface alignment and pixel-level reconstruction accuracy through geometry-aware initialization, while Reliability further uses local geometric reliability to modulate anisotropic support, effectively mitigating the smoothing of high-frequency textures caused by excessive shaping. Together with CGO’s numerical-consistency and independent controlled verification, the results indicate that FormFlow-GS can achieve a more appropriate balance among geometric consistency, pixel-level reconstruction accuracy, and local perceptual quality while maintaining efficient optimization.

AuthorsOpenReview ↗

Response to Reviewer eC7e (Q2): CGO Runtime Benefits in Robotic Mapping

FormFlow-GS reduces Mapping time on all six tested backbones: Gaussian-LIC², GS-LIVM, PVG, SGAD-SLAM, VarSplat, and MonoGS. Independent optimizer experiments further validate CGO’s efficiency advantage in frequent Gaussian parameter updates. We thank the reviewer for their interest in CGO’s practical benefits in complete robotic mapping pipelines.

Through a shared visible-Gaussian working set, CGO fuses updates of six heterogeneous attribute groups into a single CUDA Kernel, reducing repeated launches and scheduling overhead without modifying existing tracking, projection, or rendering modules.

1. Optimizer-level benefits of CGO and applicable conditions

CGO consolidates updates of six heterogeneous Gaussian attribute groups into a shared visible working set, reducing repeated launches and scheduling overhead through a single CUDA Kernel. We performed matched timing comparisons with SelectiveAdam and AdamW-GS at different Gaussian scales and visibility rates.

Table 1: Optimizer execution time at different Gaussian scales and visibility rates

Gaussian countVisibility rateCGO ↓ (μs)SelectiveAdam ↓ (μs)AdamW-GS ↓ (μs)CGO speedup
10k10%4.32314.10119.2353.26×
10k50%5.40815.18721.3302.81×
10k100%6.29816.19821.9882.57×
100k10%23.35632.20145.4461.38×
100k50%59.31970.41778.2641.19×
100k100%181.429185.881187.3921.02×
1M10%422.222424.527439.1001.01×
1M50%1186.4311182.3201200.0400.997×
1M100%1788.5931796.2711810.6051.004×

Note: A speedup greater than 1 means that CGO is faster.

The results show a clear advantage for CGO in small-scale, frequent Gaussian updates. As the Gaussian count increases, parameter and momentum reads and writes increasingly dominate, and the benefit of fused execution decreases accordingly; at 1M Gaussians, performance is already essentially on par with SelectiveAdam.

2. Relationship between optimizer benefits and full Mapping benefits

We further distinguish three measurement levels: parameter updates, local mapping optimization, and full Mapping.

Measurement levelExperimental configurationMeasured time change
Composite updateGaussian-LIC²0.997→0.723 s(−27.48%)
Mapping OptimizeGaussian-LIC²6.147→5.713 s(−7.06%)
Mapping under the adaptation protocolGS-LIVM−2.30%
Mapping under the adaptation protocolSGAD-SLAM−1.99%
Mapping under the adaptation protocolVarSplat−0.62%

These results indicate that CGO can effectively reduce Gaussian parameter-update costs, while the speedup in full Mapping also depends on the optimizer’s share of the total computational workload. When tracking, projection, and rendering consume more computation time, the acceleration of the update stage accounts for a correspondingly smaller proportion of the overall runtime.

In summary, CGO demonstrates clear optimizer-level acceleration and good cross-framework applicability. FormFlow-GS improves Mapping time on all six tested Gaussian Mapping backends. Without modifying existing tracking, projection, or rendering modules, CGO provides an efficient, lightweight, and reusable parameter-update mechanism that helps reduce the computational overhead of online robotic mapping and effectively supports low-latency, real-time Gaussian Mapping.

AuthorsOpenReview ↗

Response to Reviewer eC7e (Q3): Baseline Quality and Cross-Backbone Evaluation

FormFlow-GS has been successfully adapted to six Gaussian Mapping backends and achieves improvements in rendering quality under matched experimental conditions, validating its cross-framework integrability. We thank the reviewer for their attention to low-quality baselines and evaluation validity. We further verified the experimental configurations and original records, explicitly distinguishing controlled adaptation experiments from each method’s official standard evaluation.

1. Causes of baseline quality degradation

The baseline PSNR values for GS-LIVM, MonoGS, and SGAD-SLAM are 6.175, 7.980, and 12.378 dB, respectively. All come from actual runs and can be recomputed from the original records. To verify FormFlow-GS’s cross-framework adaptability, we used the official public implementations of these methods and adapted their mapping modules to different sensor configurations and data scenarios. Differences in scene geometry, sensor observations, and illumination conditions may affect the baselines’ absolute rendering quality. For GS-LIVM and MonoGS, the experiments use externally provided camera poses; therefore, the low PSNR values are not caused by tracking-frontend divergence in this evaluation.

We further verified the specific experimental configurations:

BackboneAdaptation experiment settings
GS-LIVMThree FAST-LIVO2 sequences; independent single-image submaps, 100 optimization updates per image; fixed poses and held-out-view rendering
MonoGSSix sequences; 9 mapping views and 20 held-out views; 500 optimization updates in total; window size 4; fixed poses
SGAD-SLAMThree TUM RGB-D sequences; independent submaps, 100 optimization updates per image; held-out-view evaluation

2. Validity of cross-framework comparisons

Table 1 uses strictly matched Base and Ours+Base comparisons, maintaining identical input data, camera poses, and optimization configurations within each backbone to evaluate performance changes after integrating GeoBirth and CGO. All experiments are based on official public implementations and follow explicitly defined adaptation protocols. Thus, the paired experiments directly measure the proposed modules’ contributions to rendering quality and mapping efficiency, providing experimental evidence for FormFlow-GS’s cross-framework integrability and performance transfer.

In summary, through GeoBirth’s geometry-aware initialization and CGO’s efficient parameter updates, FormFlow-GS successfully adapts to six Gaussian Mapping backends and demonstrates rendering-performance benefits and improved mapping efficiency in matched experiments. These results reflect the complementary advantages of the two innovations and the method’s cross-framework applicability to multimodal online mapping.

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

This paper proposes FormFlow-GS, a plug-and-play framework for online multimodal Gaussian mapping. It contains two main components: GeoBirth, which uses local metric geometry and neighborhood reliability to initialize surface-aligned anisotropic Gaussians, and CGO, which coalesces multi-attribute Gaussian updates into a shared visible-Gaussian workset to reduce optimization overhead. The paper evaluates the method on multiple datasets and transfers it to several Gaussian mapping backbones.

Critical strengths
  1. The cross-backbone evaluation is a strong point. The authors test the method on Gaussian-LIC2, GS-LIVM, PVG, SGAD-SLAM, VarSplat and MonoGS, rather than only showing results on one backbone.

  2. The 2×2 decomposition is useful. It makes the effects of GeoBirth and CGO relatively easy to understand. The supplementary experiments on birth-time geometry and final surface quality also make the GeoBirth story more complete.

  3. The efficiency improvement is practical. CGO clearly reduces optimizer-level GPU cost and kernel launches. For an online system, this engineering contribution has real value.

Critical weaknesses
  1. The main concern is novelty. GeoBirth is quite close in spirit to previous geometry-aware Gaussian initialization methods. For example, GS-LIVM already uses local covariance to initialize Gaussian geometry, while RMGS-SLAM uses voxel-PCA geometric priors for Gaussian initialization. Surface-aligned Gaussian representations are also not new. ([Open Access CVF][1]) The reliability-controlled shaping is a nice detail, but currently I am not fully convinced that it makes a sufficiently new technical contribution for ICLR.

  2. The novelty of CGO is also a little weak. Visibility-aware sparse/fused Gaussian optimization already exists. For example, gsplat provides SelectiveAdam with visibility-controlled updates and fused CUDA execution, and ICLR 2026 AdamW-GS explicitly studies sparse Gaussian optimization. ([GitHub][2]) CGO further fuses heterogeneous attribute groups into one dispatch, but the paper needs a more direct comparison with these close alternatives. Otherwise, it looks more like a good CUDA/system optimization than a new optimization method.

  3. The quality gain is not fully consistent. On Gaussian-LIC2, PSNR and SSIM improve, but the averaged LPIPS becomes worse, from 0.2390 to 0.2501. In the per-sequence results, LPIPS also becomes worse on many R3LIVE sequences and several other scenes. This makes the statement of “consistent rendering improvement” a little too strong.

  4. The end-to-end efficiency gain is sometimes small compared with the optimizer-level gain. For example, the optimizer GPU time can decrease by around 80–90%, while total mapping time only improves by around 1–2% on some backbones. This is understandable, but then the practical importance of CGO depends strongly on the backbone. A more detailed runtime breakdown would help.

  5. Some experimental controls are still missing. GeoBirth contains many manually chosen parameters, such as (K=16), neighborhood thresholds, confidence weights 0.4/0.4/0.2, activation thresholds, and the final anisotropy ratio. I did not find a systematic sensitivity analysis. Also, most reported improvements are single numbers without repeated runs, standard deviation, or confidence intervals. For several gains around 0.1–0.3 dB PSNR, this matters.

Questions for authors
  1. Can the authors provide a matched comparison between GeoBirth and the covariance/PCA initialization used in GS-LIVM or RMGS-SLAM? This seems necessary to demonstrate the real novelty of GeoBirth.

  2. Is CGO mathematically identical to the native Adam optimizer, including bias correction, optimizer step counting, and moment updates? If yes, why can CGO-only variants noticeably change PSNR/LPIPS?

  3. Can the authors compare CGO against SelectiveAdam/Sparse Adam/AdamW-GS or another optimized Gaussian optimizer, instead of only comparing against the native implementation?

3 comments
AuthorsOpenReview ↗

Response to Reviewer faUq (Q1): GeoBirth Novelty and Matched GPR/PCA Comparisons

By combining local geometric-validity assessment with Reliability-adaptive anisotropic shaping, GeoBirth outperforms GS-LIVM GPR and fixed PCA initialization on all five mean geometric metrics across 10 matched validation scenes, validating its independent value as a geometric innovation. We thank the reviewer for the suggestion to perform matched comparisons.

1. Distinctive innovation of GeoBirth and matched experiments

Unlike conventional PCA initialization, which determines surface orientation solely from local covariance, GeoBirth additionally uses local planarity, compactness, and neighborhood support to assess geometric reliability and adaptively control the anisotropic scales of Gaussians, improving geometric consistency between newly created Gaussians and the true surface.

We use the official GS-LIVM GPR initialization and PCA Fixed as matched baselines. Given that RMGS-SLAM adopts a cascaded initialization mechanism combining voxel-PCA, geometric reliability, and feed-forward prediction, this control design more directly isolates GeoBirth’s independent contribution relative to GPR and fixed PCA shaping.

The experiments strictly match Gaussian birth candidates, geometric support, birth-admission decisions, and optimization settings. They cover 10 validation scenes from M2DGR, R3LIVE, and FAST-LIVO2, with 3 paired random seeds per scene.

Table A: Mean geometric metrics over 10 validation scenes and local validation results for M2DGR room02

Evaluation scopeGeometric metricGS-LIVM GPRPCA FixedGeoBirth
10-scene meanAccuracy ↓ (cm)13.15512.68512.653
10-scene meanCompleteness ↓ (cm)94.67085.88783.423
10-scene meanChamfer ↓ (cm)53.91349.27048.054
10-scene meanF@2cm ↑ (%)34.30135.61935.738
10-scene meanNormal Error ↓ (°)30.70229.55329.386
M2DGR room02Chamfer ↓ (cm)5.19±0.504.85±0.174.57±0.13
M2DGR room02F@2cm ↑ (%)44.59±0.4845.72±0.1545.83±0.22
M2DGR room02Normal Error ↓ (°)46.63±0.2046.23±0.1245.68±0.39

Note: The first five rows are equally weighted means over the 10 scenes; the last three rows are the three-random-seed mean±standard deviation for M2DGR room02. Bold indicates the best result.

2. Analysis of the advantages in geometric reconstruction

GeoBirth achieves the best results on all five mean geometric metrics, reducing Chamfer Distance by 10.87% relative to GPR and by 2.47% relative to PCA Fixed. On M2DGR room02, Chamfer further decreases from 4.851 cm for PCA Fixed to 4.572 cm (−5.75%), while F@2cm and normal error also improve. The accompanying LiDAR reference-geometry overlays, paired distance-error maps, and normal sections further illustrate GeoBirth’s improvement in local surface geometric displacement.

In summary, GeoBirth’s core innovation is to extend conventional static PCA surface alignment into a reliability-aware, adaptive Gaussian geometric-shaping mechanism. Strictly matched multi-scene quantitative experiments and local geometric visualizations jointly demonstrate that this design can effectively improve surface geometric-alignment accuracy and three-dimensional reconstruction quality, establishing GeoBirth’s independent contribution and practical advantages over existing GPR/PCA initialization methods.

AuthorsOpenReview ↗

Response to Reviewer faUq (Q2): CGO–Adam Equivalence and Rendering Differences

CGO achieves efficient fused multi-attribute execution while preserving native parameter-update semantics: 10,692 updates are bitwise identical, and independent controlled experiments found no stable effect on rendering quality. We thank the reviewer for their attention to CGO’s numerical correctness and rendering differences. We systematically verified these issues through source-code auditing, bitwise-consistency verification, and supplementary experiments.

1. Mathematical equivalence between CGO and native Adam

The source-code audit confirms that the tested native Gaussian-LIC² implementation uses visibility-gated Adam without bias correction, and the optimizer step count does not enter its parameter-update formula. CGO preserves the same learning rates, first- and second-moment recurrences, numerical-stability term, and visibility rules. The parameters and momenta of invisible Gaussians remain unchanged. Under identical parameters, gradients, and optimizer states, all 10,692 parameter and momentum updates are bitwise identical, verifying consistency between CGO and the tested native Adam update semantics.

2. Source of CGO-only image-quality differences

To localize the differences, we conducted controlled experiments on 4 real sequences, with 3 paired seeds and 72 training trajectories. The results show that, across 84 trajectory comparisons, the first numerical difference consistently occurs during Backward gradient accumulation at step 1, before the Adam update.

Further fixed-input replay shows that atomicAdd floating-point gradient accumulation in Backward is sensitive to parallel execution order and can produce small rounding variations that propagate through subsequent optimization. This provides a reasonable explanation for the historical E00/E01 rendering differences.

After further fixing the Backward configuration and switching only the optimizer, we obtained the following results:

Table 1: PSNR effect of CGO with the Backward configuration fixed

SequenceCGO − Native PSNR (dB)95% CI
M2DGR room01+0.0000027[−0.0000696, +0.0000749]
R3LIVE degenerate 00−0.0001336[−0.0006528, +0.0003857]
FAST-LIVO2 CBD Building 01−0.0004337[−0.0025258, +0.0016585]
R3LIVE campus 00+0.00000225[−0.0002128, +0.0002867]

The PSNR confidence intervals for all four sequences include zero, and optimizer main-effect tests for SSIM and standard LPIPS also found no stable changes. This indicates that, under the tested conditions, CGO parameter updates themselves show no detectable independent effect on image quality; the historical differences are more consistent with an explanation based on upstream Backward floating-point accumulation and execution-configuration changes.

In summary, CGO preserves mathematical update semantics consistent with the tested native Adam. Its parameter and momentum updates have passed bitwise-consistency verification, and the historical image-quality differences have also been localized to Backward, before the optimizer. CGO’s core innovation is to reduce optimization-execution overhead through a shared visible-Gaussian working set and fused multi-attribute updates while maintaining numerical correctness, providing an efficient, reliable, and reusable optimization mechanism for different Gaussian Mapping backends.

AuthorsOpenReview ↗

Response to Reviewer faUq (Q3): Comparison with Optimized Gaussian Optimizers

CGO achieves the fastest execution time in eight of the nine optimizer comparisons and demonstrates substantial advantages in small- and medium-scale Gaussian updates, validating the independent efficiency contribution of its fused multi-attribute update mechanism. We thank the reviewer for the suggestion. We added direct comparisons with gsplat SelectiveAdam and AdamW-GS.

1. Comparison methods and experimental settings

Beyond selective updates that simply skip invisible Gaussians, CGO organizes six heterogeneous attribute groups, totaling 59 dimensions, into a shared visible working set and fuses their execution into a single CUDA Kernel to reduce repeated launches and scheduling overhead. We use SelectiveAdam and AdamW-GS, which support visibility-based selective updates, as the main controls. PyTorch SparseAdam primarily handles sparse-format gradients, which differ from the explicit visibility masks and dense Gaussian attribute storage used in this experiment.

The experiments use matched FP32 parameters, gradients, and visibility masks across three Gaussian scales and different visibility rates. The supplementary experiments include 60 configurations, 420 repeated measurements, and Profiler data for 42,000 steps.

Table 2. Optimizer execution time at different Gaussian counts and visibility rates (μs; lower is better)

Gaussian countVisibility rateCGO ↓SelectiveAdam ↓AdamW-GS ↓CGO speedup
10k10%4.32314.10119.2353.26×
10k50%5.40815.18721.3302.81×
10k100%6.29816.19821.9882.57×
100k10%23.35632.20145.4461.38×
100k50%59.31970.41778.2641.19×
100k100%181.429185.881187.3921.02×
1M10%422.222424.527439.1001.01×
1M50%1186.4311182.3201200.0400.997×
1M100%1788.5931796.2711810.6051.004×
Kernel launches—167—

2. Experimental results

CGO achieves a 2.57–3.26× speedup at 10k Gaussians and a 4.45× speedup relative to AdamW-GS at 10% visibility; at 100k Gaussians, it retains a 1.02–1.38× speedup. At 1M Gaussians, CGO’s execution time is essentially on par with SelectiveAdam.

These results indicate that CGO is particularly suitable for frequent Gaussian parameter updates at small and medium scales and low visibility rates, where its single-Kernel fused design can effectively reduce repeated execution overhead.

In summary, CGO retains the advantages of visibility-based selective updates and further improves optimization efficiency through a shared working set and fused multi-attribute execution. Direct comparisons with existing optimized Gaussian Optimizers validate its independent computational advantages, providing a general, efficient, and easily integrated parameter-update mechanism for different Gaussian Mapping backends.

Reviewer_mSX5OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FormFlow-GS proposes two plug-in modules for online Gaussian mapping. GeoBirth uses local LiDAR or depth geometry and a reliability score to initialize Gaussian orientation and shape. CGO combines visibility-masked Adam updates across attribute groups into one kernel to reduce optimization overhead. The paper evaluates Gaussian-LIC2 on 12 sequences and transfers the modules to five additional backbones, using factorial experiments to separate their effects. On Gaussian-LIC2, the combination improves PSNR by 0.35 dB and reduces mapping time by 11.75%; geometry evaluations on three sequences also report improvements across all five surface metrics.

Critical strengths
  1. A demonstrated execution improvement over an optimized baseline. With formation fixed, CGO reduces mapping time across all six backbones (Tables 2 and 10). On Gaussian-LIC2, it reduces time from 135.41 to 124.72 s, and the full combination reaches 119.50 s. Gaussian-LIC2 already uses per-group visibility-masked Adam, so the improvement is measured against an already efficient sparse optimizer. This is the clearest practical contribution.

  2. GeoBirth improves reconstructed geometry. On all three evaluated sequences, it improves accuracy, completeness, Chamfer distance, F@2cm and normal error (Table 7, Figures 7-8). The factorial experiments separately evaluate formation and execution. Together, these provide substantive evidence that GeoBirth contributes beyond a timing improvement and are the strongest argument for accepting the paper despite modest rendering gains.

Critical weaknesses
  1. The practical quality gains are modest or difficult to interpret. Gaussian-LIC2, PVG and VarSplat gain only 0.10-0.35 dB; Gaussian-LIC2's reported LPIPS increases from 0.2390 to 0.2501, without reported run-to-run variability. Three other backbone configurations render at only 6-12 dB (Table 1). Their matched comparisons support integration and efficiency transfer, but offer limited evidence of useful rendering-quality transfer. This leaves the overall practical advance less compelling than the breadth of the evaluation suggests.

  2. The distinctive benefit of reliability modulation remains unisolated. Appendix A.2 explains its behavior, and the GS-LIVM comparison already improves on a covariance-based initializer (Table 10: 6.175->6.408 dB with GeoBirth alone). However, that comparison changes the complete initializer. A comparison holding the PCA surface frame and support construction fixed would establish whether reliability modulation itself improves the geometry metrics. This matters to my recommendation because it would establish the added research value of GeoBirth's central mechanism.

Questions for authors

Clarification (Weakness 1): Which operating settings explain the 6-12 dB baseline reconstructions, and what variability is observed in Gaussian-LIC2's modest PSNR/LPIPS changes? This would help establish the practical significance and stability of the reported rendering improvements. Additional evidence (Weakness 2): Does reliability modulation improve the Table 7 geometry metrics when the same PCA surface frame and support construction are retained? This would distinguish the contribution of reliability modulation from the underlying geometric initialization.

2 comments
AuthorsOpenReview ↗

Response to Reviewer mSX5 (Weakness 1): Baseline Conditions and Rendering Stability

FormFlow-GS has been successfully adapted to six Gaussian Mapping backends and improves PSNR on 11 of the 12 test sequences for Gaussian-LIC². Supplementary three-random-seed experiments further validate the stability of rendering-quality improvements, demonstrating the method’s cross-framework applicability and reconstruction-performance advantages. We thank the reviewer for their attention to baseline operating conditions and experimental reproducibility. We further verified the cross-framework adaptation protocols and original evaluation records and added statistics from multiple independent runs.

1. Cross-backbone adaptation and baseline operating conditions

The baseline PSNR values for GS-LIVM, MonoGS, and SGAD-SLAM in Table 1 are 6.175, 7.980, and 12.378 dB, respectively. All come from actual runs and can be recomputed from the original records.

Using the official public implementation of each method, we adapt FormFlow-GS to different sensor configurations and data scenarios. The specific settings are as follows:

BackboneAdaptation experiment settings
GS-LIVMThree FAST-LIVO2 sequences; independent single-view submaps, 100 optimization updates per image; fixed camera poses and held-out-view evaluation
MonoGSSix sequences from M2DGR, R3LIVE, and FAST-LIVO2; 9 mapping views and 20 held-out views; 500 optimization updates; window size 4; fixed camera poses
SGAD-SLAMThree TUM RGB-D sequences; independent submaps, 100 optimization updates per image; held-out-view evaluation

Differences in scene geometry, sensor observations, and illumination conditions may affect the baselines’ absolute rendering performance. For GS-LIVM and MonoGS, the experiments use external camera poses; therefore, the low PSNR values are not caused by tracking-frontend divergence in the tested pipeline.

Within each backbone, Base and Ours+Base use identical inputs, poses, optimization configurations, and evaluation protocols, ensuring that performance changes fairly reflect the effects of integrating FormFlow-GS’s modules.

2. Rendering-quality improvements and statistical stability

Appendix Table 12 of the original manuscript reports the complete Gaussian-LIC² rendering results on 12 sequences, with PSNR improvements on 11/12 sequences, demonstrating cross-scene consistency in pixel-level reconstruction benefits.

To further verify run-to-run stability, we conducted controlled E00/E11 replays in three representative scenes using 3 paired random seeds. E00 is the native configuration, while E11 enables both GeoBirth and CGO. Each run includes 297 aligned frames and 1,770 optimization updates.

Table 1: Three-random-seed rendering results for E00 and E11 on Gaussian-LIC² (Mean ± SD, n=3n=3)

SceneMetricE00 Native Mean ± SDE11 FormFlow-GS Mean ± SD
M2DGR room03PSNR (dB) ↑17.380 ± 0.006917.649 ± 0.0077
LPIPS ↓0.3340 ± 0.00040.3438 ± 0.0004
R3LIVE degenerate 01PSNR (dB) ↑22.940 ± 0.027923.356 ± 0.0468
LPIPS ↓0.1590 ± 0.00120.1767 ± 0.0012
FAST-LIVO2 Bright Screen WallPSNR (dB) ↑24.758 ± 0.009525.543 ± 0.0290
LPIPS ↓0.1512 ± 0.00070.1440 ± 0.0010

Note: Mean ± SD denotes the mean±sample standard deviation of three paired independent runs (n=3n=3);

E11 achieves higher PSNR in all three scenes, with small standard deviations across runs, indicating that the joint GeoBirth and CGO configuration can produce repeatable pixel-level reconstruction improvements. Changes in LPIPS across scenes further reflect the tradeoff between pixel accuracy and perceptual detail. In addition, Table 2 of the original manuscript separates the independent contributions of GeoBirth and CGO through a 2×2 ablation, while Appendix Table 8 provides geometric evaluation evidence for the final three-dimensional surface reconstruction.

In summary, FormFlow-GS demonstrates good cross-framework adaptability across six Gaussian Mapping backends, and its pixel-level reconstruction advantages and run-to-run stability are validated through rendering evaluations on 12 sequences and three-random-seed experiments. GeoBirth’s geometry-aware initialization and CGO’s efficient fused optimization complement one another, providing a general solution that balances reconstruction quality and computational efficiency for online Gaussian Mapping under different sensor configurations.

AuthorsOpenReview ↗

Response to Reviewer mSX5 (Weakness 2): Independent Geometric Contribution of Reliability

Under identical PCA geometric initialization, Reliability reduces Chamfer Distance by 7.46% and 3.17% in the two test scenes, respectively, while improving surface completeness and effective support coverage, validating the independent geometric contribution of reliability-adaptive shaping. We thank the reviewer for the suggestion concerning Reliability’s independent role.

1. Independent Reliability ablation experiments

We fix PCA surface frames, geometric support, Gaussian birth candidates, and optimization settings, changing only the shaping-control strategy, and establish three controls: fixed PCA (B), mean confidence (B2), and shuffled confidence (D). B2 applies the scene-wide mean reliability activation coefficient uniformly to all Gaussians to control average shaping strength. D randomly shuffles the per-Gaussian reliability activation coefficients, preserving their numerical distribution while destroying the correspondence between reliability and spatial location. Through these controls, we further distinguish the contribution of shaping strength itself from that of local spatial reliability information to three-dimensional reconstruction quality.

Table 1: Reliability geometry ablation under the same PCA initialization

Development sceneB / B2 / D Chamfer ↓ (cm)Reliability Chamfer ↓ (cm)Improvement over BCompleteness ↓(B→F)Support ↑(B→F)
M2DGR room033.420 / 3.248 / 3.1583.1657.46%3.576→3.16029.394→33.683%
FAST-LIVO2 Bright Screen Wall2.494 / 2.880 / 2.8122.4153.17%2.238→2.07136.991→38.032%

Note: Both Completeness and Support denote B → Reliability; bold marks the best Chamfer values within each group and the relative improvement.

2. Geometric reconstruction benefits

Reliability outperforms both fixed PCA (B) and mean confidence (B2) in the two scenes, reducing Chamfer and Completeness while increasing Support. In particular, on Bright Screen Wall, Reliability achieves the lowest Chamfer Distance (2.415 cm), outperforming B, B2, and D, demonstrating the advantage of using local geometric reliability for spatial shaping.

In summary, while preserving PCA’s surface-alignment capability, Reliability further improves three-dimensional reconstruction accuracy, surface completeness, and effective coverage by adaptively modulating the anisotropic support of Gaussians. Matched ablation results validate Reliability’s independent contribution relative to fixed PCA initialization, further demonstrating the innovative value of GeoBirth’s reliability-aware geometric-initialization mechanism.

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

Related papers

Discussion (0)

Sign in to comment.