acceptodds
Withdrawn from ICLR 2027

Necromantic Gaussian: Raising the Retention Confidence for Continuous Level of Detail in 3D Gaussian Splatting

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Level of detail (LoD) for 3D Gaussian Splatting trades rendering cost against fidelity, but reducing a trained field requires deciding which primitives to retain. Heuristics based on size, distance, or opacity can discard essential coverage, while hard selection removes the reconstruction feedback that could reverse a bad decision. We introduce Necromantic Gaussian, which learns retention confidence from the rendering loss itself. Built on Octree-GS, a lightweight router predicts per-anchor confidence from base and dedicated side features. Rendering supervision teaches which anchors are kept, while a direct count objective calibrates how many decoded Gaussians are selected at a requested budget. An extended rasterizer backward pass evaluates local reconstruction sensitivity at zero opacity, allowing rejected gates to receive image evidence without contributing to the hard forward image. Scale compensation and joint optimization of the base representation help the retained subset preserve coverage. The resulting confidence ordering lets a single checkpoint serve budgets specified at inference. We evaluate its quality–workload trade-offs on Tanks&Temples and Mip-NeRF 360, including controlled mechanism ablations.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
gaussian splatting, computer vision

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_jZjMOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The submission proposes "Necromantic Gaussian," a continuous Level-of-Detail (LoD) framework built on top of Octree-GS. The method formulates runtime primitive simplification as a learned retention confidence ordering. An anchor-level router MLP takes base features and learned side features to predict a retention logit and slot-wise scale compensation. To address the gradient vanishing problem in hard primitive selection, the authors introduce a zero-opacity analytic backward pass where rejected primitives receive local alpha-compositing gradients without contributing to the forward rendered image. Additionally, a direct decoded-Gaussian count loss calibrates the confidence distribution against a requested workload budget. A single trained model is evaluated across continuous budget fractions (100%, 50%, 25%) on Mip-NeRF 360, Tanks&Temples, MatrixCity, and vanilla 3DGS.

Critical strengths

S1: Formulating continuous LoD through a unified, learned confidence ordering that supports runtime budget slicing without retraining or multi-checkpoint storage addresses an important deployment challenge in explicit radiance field rendering. S2: The conceptual decomposition between reconstruction feedback (determining which primitives preserve visual fidelity) and the straight-through decoded count penalty (controlling how many total primitives are admitted) offers a structured and theoretically clean perspective on budget-constrained primitive pruning.

Critical weaknesses

W1: The central methodological premise is empirically contradicted by the ablation study in Table 2. The primary technical selling point of this paper is the zero-opacity analytic backward pass (the "necromantic" gradient recovery mechanism), designed to prevent rejected primitives from getting trapped in non-recoverable deadlocks. However, Table 2 reveals that disabling this mechanism ("No analytic backward") leads to negligible differences on Mip-NeRF 360 (-0.055 dB at 50% budget, -0.101 dB at 25% budget) and actually improves PSNR on Tanks&Temples by +0.055 dB at 50% budget and +0.044 dB at 25% budget. Furthermore, removing the scale compensation module at 25% budget increases PSNR on Mip-NeRF 360 by +0.095 dB and on Tanks&Temples by +0.028 dB; removing warmup at 25% budget similarly yields performance increases (+0.022 dB and +0.014 dB). When core components exhibit negative or marginal contributions in realistic multi-view reconstructions, the claim that they are necessary to resolve pruning deadlocks collapses.

W2: Disconnect between Gaussian reduction and real-world rendering throughput in Table 5. The fundamental motivation of an LoD framework is to translate primitive reduction into substantial inference acceleration. In Table 5 on Mip-NeRF 360, reducing the budget from s=1s=1 to s=4s=4 cuts the active Gaussian count from 437k to 112k—a 74.2% reduction in primitives. Yet, the rendering speed only scales from 299.73 FPS to 352.91 FPS—a meager 17.7% gain. The authors do not provide a profiling breakdown to explain why cutting nearly three-quarters of the primitive workload yields such marginal runtime benefits, suggesting that per-anchor confidence computation, dynamic sorting, and threshold masking introduce substantial pipeline overhead that negates rasterization gains.

W3: Performance degradation over the underlying base model and inferior performance against existing baselines in Table 1. At 100% full capacity (s=1s=1), the proposed model achieves 27.709 dB on Mip-NeRF 360, underperforming its own foundational base architecture (Octree-GS achieves 28.019 dB). This indicates that the joint training of the router and side features penalizes the capacity of the base representation. At restricted budgets (50% and 25%), Matryoshka-GS outperforms the proposed method by 1.0 to 1.5 dB across both Mip-NeRF 360 and Tanks&Temples. The justification provided in the text regarding differing primitive counts does not alter the fact that the proposed approach delivers an inferior rate-distortion Pareto frontier.

W4: Unsubstantiated claims regarding generalization and representation independence in Tables 4 and 6. The authors claim in Section 4.5 that their method scales to large urban scenes (MatrixCity, Table 4) and in Section 4.7 that it is representation-agnostic (Vanilla 3DGS, Table 6). However, both tables solely report isolated metrics of the proposed method across scales (s=1,2,4,8s=1, 2, 4, 8) with absolutely zero baseline comparisons. Without comparing MatrixCity results against native Octree-GS or CityGS, or comparing Vanilla 3DGS results against standard opacity/scale pruning or random dropout baselines, these tables offer no comparative scientific value.

W5: Deficient manuscript polish and careless preparation. The paper contains blatant uncompiled syntax errors in the main text; specifically, on line 047 in the Introduction, an unresolved citation appears as ... (Ren et al., 2024; Kerbl et al., 2024; Fan et al., 2024; Liu et al., 2025; ? ; Guo et al., 2026). Additionally, the qualitative comparisons in Figure 4 contain low-resolution crops without magnified insets or difference error maps, making visual verification impossible. The submission metadata (e.g., the informal TL;DR "a type of lod(level of detail) for gaussian splatting") further reflects rushed preparation and a lack of academic rigor.

Questions for authors

Q1: In Table 2, why does removing the zero-opacity analytic backward pass improve PSNR on Tanks&Temples (+0.055 dB at 50% and +0.044 dB at 25%), and why does removing scale compensation yield a +0.095 dB boost on Mip-NeRF 360 at 25%? How do the authors reconcile these empirical results with the claim that these components are indispensable for recovering from bad pruning decisions and preserving coverage?

Q2: What is the detailed runtime profile of Table 5? Specifically, what fraction of latency is consumed by anchor router forward passes, confidence evaluation, and filtering, versus the CUDA rasterization stage? Why does a 74.2% drop in rendered Gaussians result in only a 17.7% FPS improvement?

Q3: Why are Tables 4 and 6 reported without baseline methods? What are the quantitative PSNR/SSIM/LPIPS numbers of standard Octree-GS on MatrixCity and simple opacity/magnitude thresholding on Vanilla 3DGS under identical Gaussian budgets?

Q4: Given that the proposed model incurs an initial quality drop at full capacity compared to base Octree-GS (27.709 dB vs. 28.019 dB on Mip-NeRF 360), what specific regularization or optimization dynamics cause the base field to degrade during joint routing supervision?

Reviewer_VAj5OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Necromantic Gaussian learns anchor-level retention confidence on top of Octree-GS. Reconstruction supervision learns a shared ordering, while a decoded-Gaussian count objective calibrates workload. A modified rasterizer supplies gradients to zero-opacity rejected candidates, and scale compensation adapts retained coverage. Inference uses a workload-aware prefix or threshold gating to expose multiple budgets from one checkpoint.

Critical strengths
  1. Explicit workload calibration. Counting decoded Gaussian slots rather than anchors addresses the fact that equal numbers of anchors can produce different workloads.
  2. Competitive rendering at reduced workloads. Table 1 reports better image quality with fewer effective Gaussians than the listed reduced-budget CLoD-GS and FlexGaussian operating points. The same checkpoint supports multiple budgets. These are meaningful system-level results, although effective Gaussian counts alone do not establish runtime or memory superiority.
  3. Several informative controls. The paper distinguishes per-view selection from global subsets and reports analytic-backward, warmup, compensation, and threshold-versus-prefix ablations rather than relying only on headline results.
Critical weaknesses
  1. The practical value of gradient recovery remains insufficiently established. In Table 2, removing analytic backward changes PSNR by −0.055/−0.101 dB on Mip-NeRF360 and +0.055/+0.044 dB on Tanks and Temples. The authors explicitly motivate this mechanism by recoverability rather than final PSNR, which is a reasonable distinction. However, evidence of recovery frequency, affected image regions, or robustness under difficult training conditions is needed to establish its practical importance. Small or mixed ablation effects do not invalidate the framework’s positive overall results.
  2. The contributions of the training components are not fully separated. The training top-k variant changes both selection and count calibration and is explicitly an alternative training strategy. Its substantial degradation supports the complete training recipe over that alternative, but cannot isolate the contribution of gradient recovery or decoded-count supervision. More controlled ablations would clarify which mechanisms account for the reported quality–workload advantage. Scale compensation also has small, mixed effects and should be positioned as a secondary refinement.
  3. Efficiency evidence and metric consistency need clarification. Retaining approximately one quarter of Gaussians increases FPS from 299.73 to 352.91 on Mip-NeRF360 and from 274.28 to 330.23 on Tanks and Temples. These are useful gains, but matched-quality end-to-end timing against alternatives and a stage-wise runtime breakdown are needed to establish the efficiency advantage; linear scaling with counts is not expected. Separately, at full budget Table 1 reports LPIPS 0.2107/0.1937, whereas Table 5 reports 0.175/0.161, with matching PSNR and approximately matching counts. Please explain any metric or evaluation differences.
Questions for authors
  1. How often do rejected anchors recover during real-scene training, and does this improve particular regions, robustness, or convergence? Can the controlled recovery experiment be connected to benefits in the evaluated scenes?
  2. Can confidence learning, decoded-count supervision, and gradient recovery be isolated while holding the remaining routing and training settings fixed? Which component explains the gain over the training top-k alternative?
  3. How do threshold and prefix modes compare with competing methods at matched quality in end-to-end time, and how much time is spent in confidence prediction, decoding, selection, and rasterization?
Reviewer_zvWeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method that learns which Gaussians are useful for a given scale, rather than relying on heuristics based on scale or opacity. Based on the anchor setup used in ScaffoldGS and OctreeGS, the proposed method additionally learns soft-confidence values that are hard-masked to select Gaussians using a passed-in scale parameter. Additionally, they correct for the loss introduced by the confidence via scale compensation and add a loss term that promotes the number of Gaussians to follow the selected virtual scaling term.

Critical strengths

The paper proposes an interesting approach to performing LoD, in which selection is learned rather than based on heuristics. While similar in nature to Compact3DGS, adding LoD support is a useful addition.

The evaluation covers a broad range of existing methods and shows that it could be applied beyond anchor-based methods.

Critical weaknesses

The first major weakness of the paper is its evaluation. While it covers several related works, it is limited to two datasets, does not compare against OctreeGS for lower scales, on which it is based, and the ablation does not motivate each component, as the metric changes are within the noise of multiple runs for the same method. Additionally, there is no FPS evaluation against other methods. Compact3DGS also proposes a different and simpler approach to workload correction, for which no comparison is provided.

Second, the paper, in its current form, is not reproducible because various parameters are not explained (E1: kappa, E4: l_epsilon, E5: l_i, l_i^0,t_i,.. E6: w(s),...). The paper should clarify and explain what the parameters are and what they represent. On L163-164, the shared decay trend appears in the text without further explanation.

Questions for authors

Are the results of the ablation averaged over multiple runs? If so, could this be reported in the paper and the table caption, including standard deviation? If not, the ablation does not convincingly establish the inclusion of the parameters.

Does the Compact3DGS workload correction strategy, which averages the activated soft confidence, work instead of the proposed strategy in the paper? How do they compare?

How does FPS compare against existing works? Given that the MLP must be queried per anchor, a per-Gaussian correction must be computed.

Additional feedback:

Missing references: L47-48, L97-98 Images in Figure 4 are quite small. The margins around percentages are large, so the paper could reduce them to make room for larger images. In Table 2, the top and bottom margins appear very small, and the table is not referenced in the text. References to tables and sections are not capitalized, even at the start of a sentence.

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

Related papers

Discussion (0)

Sign in to comment.