acceptodds
Under review as a conference paper at ICLR 2027

LocusGS: Spatially Grounded Tokens for Feed-Forward 3D Gaussian Splatting

Abstract

Recent query-based feed-forward 3DGS methods represent a scene using learnable queries, each aggregating multi-view evidence and decoding a group of Gaussians. Ideally, different queries should specialize in coherent local regions of the scene. However, we observe that Gaussians decoded from the same query often scatter across distant scene regions, resulting in weak query-level spatial coherence and poor alignment with the scene structure. We attribute this behavior to the purely latent representation of existing Gaussian queries. To address this limitation, we introduce LocusGS, which augments each Gaussian query with a 3D anchor state consisting of a center and a support radius. The anchor state is progressively refined across decoder layers and is used throughout query interaction, multi-view feature aggregation, and Gaussian generation. Specifically, an anchor-to-ray geometric bias guides each query toward spatially relevant image observations, while anchor-centered decoding organizes its Gaussians within a local region. Experiments on novel view synthesis benchmarks show that LocusGS improves rendering quality over query-based Gaussian token baselines under the same Gaussian budget. Further analysis shows that the learned anchors form coherent spatial layouts and lead to more structured Gaussian distributions, demonstrating that explicit anchor states improve the spatial organization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3DGS, 3D reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_dRN8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes LocusGS, a query-based feed-forward 3DGS method that augments each Gaussian query with an explicit 3D anchor (a center and a radius). The anchor is refined across decoder layers and guides query interaction and cross-view feature aggregation. Each query's Gaussians are then decoded as local offsets around its anchor. Experiments on RE10K and DL3DV show improvements over TokenGS under matched budgets, supported by ablations and visualizations of the learned anchors.

Critical strengths

1. Consistent improvements under matched budgets

  • LocusGS improves over TokenGS on both datasets, across different token budgets and numbers of input views. This suggests practical value for query-based feed-forward 3DGS.

2. A reasonable idea, clearly presented

  • Giving each query an explicit 3D anchor is an intuitive and reasonable idea. The paper is well written and includes informative visualizations, which makes the method easy to understand and build upon.
Critical weaknesses

1. The motivating observation could be better supported

  • The method is motivated by the observation that Gaussians from one latent query "often scatter across distant scene regions" (Abstract; Sec. 2.1.1). This is illustrated with a representative token (Fig. 1).
  • However, the TokenGS paper reports that Gaussians from a single token "stay close to each other within a scene" (Sec. 4.5, Fig. 10). It would help to reconcile the two observations.
  • The quantitative analysis (Sec. 3.3.3, Table 3) is helpful, but it may not fully support the claim:
    • C_centroid is a mean distance pooled over all tokens and Gaussians in a scene. A small fraction of far-away, possibly near-transparent, Gaussians could therefore yield a large value even if most tokens are compact. As a result, the metric does not show how often tokens are dispersed.
    • The lower dispersion of LocusGS is largely expected from its anchor-centered decoding (Eq. 9). The comparison alone therefore does not show that dispersion affects rendering quality.
  • Why it matters: This observation is the main motivation for the anchor design. Without stronger support, it is unclear what explains the reported gains.

2. The main comparison may not isolate the effect of spatial grounding

  • Table 5 shows that each anchor component matters within LocusGS. However, it does not include TokenGS, so it does not quantify the gain over the baseline.
  • LocusGS also adds intermediate-layer supervision, which alone accounts for +0.74 dB (Table 4). This is of the same order as the reported gains over TokenGS.
  • The TokenGS baseline does not appear to use this supervision. For example, the RE10K TokenGS results in Table 2 are identical to Table 1 of the TokenGS paper, and the TokenGS training objective (its Sec. 3.1) has no intermediate supervision.
  • Why it matters: The ablations support the individual components. However, part of the improvement over TokenGS, which is the paper's main claim, may come from the added supervision. It therefore remains unclear how much of the gain is due to spatial grounding.
Questions for authors
  1. Novelty and positioning.
    • The closest design I am aware of is UniQueR (ECCV'26). It also uses learnable queries with explicit 3D positions to decode Gaussians for scene-level feed-forward reconstruction. Could you clarify the key differences from it? Given its timing, I do not count its omission against the paper.
    • Object-level methods such as UniGS (ICCV'25) and LeanGaussian (CVPR'25) also treat Gaussians as queries with explicit 3D positions. A brief discussion would help.
  2. Motivating observation (Weakness 1).
    • Could you clarify whether the TokenGS used in Figs. 1 and 5 and Table 3 is the official checkpoint or a re-trained one? Fig. 9 and Table 1 suggest the latter. If it was re-trained, does the official checkpoint show similar dispersion?
    • Could you report the distribution of per-token dispersion, weighted by opacity (e.g., the fraction of widely dispersed tokens)? If most tokens turn out to be compact, this would also reconcile Fig. 1 with TokenGS Sec. 4.5.
  3. Supervision (Weakness 2). Was TokenGS trained with the same intermediate supervision? Results for TokenGS trained with this supervision, in the Table 5 setting, would largely resolve this concern.
  4. Robustness and transfer. The anchor design relies on exact camera rays and on random initial anchors shared across scenes. Could you:
    • evaluate robustness to pose noise using the TokenGS protocol (its Sec. 4.3);
    • evaluate cross-dataset transfer (e.g., RE10K→ACID);
    • report how far the anchors move from their initialization across layers?

[1] UniQueR: Unified Query-based Feedforward 3D Reconstruction. ECCV 2026.

[2] UniGS: Modeling Unitary 3D Gaussians for Novel View Synthesis from Sparse-view Images. ICCV 2025.

[3] LeanGaussian: Breaking Pixel or Point Cloud Correspondence in Modeling 3D Gaussians. CVPR 2025.

4 comments
AuthorsOpenReview ↗

Response to Reviewer dRN8 (Part 1 / 3)

We sincerely thank the reviewer for the thorough, constructive feedback and for recognizing LocusGS’s consistent gains over TokenGS across datasets and metrics. Your comments prompted several insightful analyses, including opacity-aware Gaussian dispersion analysis, supervision-matched experiments, robustness to camera pose noise and quantitative analysis of progressive anchor refinement.

We hope these clarifications and new experiments fully address your concerns, and we would be grateful if you would consider them in your final assessment. Detailed responses to each of your questions follow below.


Q1.1: Clarify the difference between UniQueR (ECCV'26), UniGS (ICCV'25) and LeanGaussian (CVPR'25) to better position.

A1.1: Thank you for highlighting these closely related works. We agree that explicit 3D queries and Gaussian refinement have been explored previously. The key distinctions are as follows.

  1. UniQueR uses 3D query positions during feature aggregation and predicts a one-shot query deformation after its 12-layer Transformer. In contrast, LocusGS progressively updates anchor positions across decoder layers, allowing refined positions to guide subsequent geometric attention. Moreover, LocusGS learns an adaptive support radius that jointly controls anchor-to-ray attention and local Gaussian decoding.

    Interestingly, UniQueR reports a substantial PSNR drop from 20.23 to 12.11 dB with random-only initialization. In our additional experiment in Appendix Sec B.4, initializing LocusGS with Pi3-predicted point maps and FPS yields nearly identical final performance to random learnable anchors. This suggests that LocusGS is less dependent on external geometric initialization in our setting, potentially benefiting from its progressive spatial refinement.

  2. UniGS (ICCV 2025) and LeanGaussian (CVPR 2025). Both methods focus on object-level reconstruction, refining one Gaussian per query through projection-based deformable attention. In contrast, LocusGS represents a local Gaussian group per query, with a learned center and radius jointly guiding feature aggregation and Gaussian generation.

We appreciate the reviewer's suggestions and will clarify these distinctions in the revised manuscript.


Q2.1: Could you clarify whether the TokenGS used in Figs. 1 and 5 and Table 3 is the official checkpoint or a re-trained one? Fig. 9 and Table 1 suggest the latter. If it was re-trained, does the official checkpoint show similar dispersion?

A2.1: Thank you for raising this important question. We apologize for not making the checkpoint sources sufficiently clear in the paper. We would like to clarify that we prioritize using the officially released TokenGS checkpoints whenever available, to ensure a fair comparison and avoid potential differences introduced by our own re-training.

Specifically:

  1. DL3DV quantitative comparisons (e.g., Tables 1 and 3): All reported TokenGS results are obtained by evaluating the official checkpoints under our evaluation protocol. Interestingly, our evaluation produces slightly better results than those reported in the original TokenGS paper. For example, on DL3DV we obtain 23.44 dB PSNR, compared with 23.02 dB reported by TokenGS.

  2. Qualitative comparisons and 3D visualizations: All TokenGS visualizations of Gaussian distributions and spatial structures (e.g., Figs. 1 and 6), as well as qualitative rendering comparisons (e.g., Figs. 4, 8, 9, and 12), are generated using the official checkpoints. Therefore, the observed Gaussian dispersion is not an artifact of our re-training.

  3. RE10K quantitative comparisons (e.g., Table 2): Since official TokenGS checkpoints for RE10K were not available, we re-trained TokenGS following its original training setup.

  4. Training convergence (Fig. 9): To compare convergence behavior fairly, we trained both TokenGS and LocusGS under comparable settings on DL3DV and RE10K. Our re-trained TokenGS models achieve performance close to, although slightly below, the official DL3DV checkpoints, and exhibit stable training behavior. Importantly, the re-trained models show similar Gaussian spatial dispersion to the official checkpoints.

In summary, our observations regarding Gaussian dispersion hold for both the official and re-trained TokenGS models. Re-training is performed only when necessary for datasets without released checkpoints or for controlled training-convergence comparisons.

We will clarify these checkpoint sources and evaluation settings in the revised manuscript to avoid further ambiguity.


AuthorsOpenReview ↗

Response to Reviewer dRN8 (Part 2 / 3)

Q2.2: Could you report the distribution of per-token dispersion, weighted by opacity (e.g., the fraction of widely dispersed tokens)? If most tokens turn out to be compact, this would also reconcile Fig. 1 with TokenGS Sec. 4.5.

A2.2: Thank you for this insightful suggestion. We agree that opacity is important when interpreting the spatial distribution of Gaussian groups.

MethodUnweighted dispersion ↓Opacity-weighted dispersion ↓
TokenGS0.58810.1603
LocusGS0.04330.0436

These results reveal an important distinction: although TokenGS produces spatially scattered Gaussian centers, much of this dispersion is associated with low-opacity Gaussians. Thus, the compact token groups observed in TokenGS Sec.4.5 are not necessarily inconsistent with the scattered 3D centers shown in our Figure 1. Our original unweighted metric captures the distribution of all predicted centers, whereas opacity-weighted dispersion better reflects the concentration of Gaussians with greater rendering contributions.

We further illustrate this phenomenon in Figure 12, where query-colored Gaussian centers are overlaid on reference geometry, with close-up views highlighting off-surface outliers and local Gaussian organization. Together with the opacity-weighted statistics, these visualizations help clarify why scattered 3D centers can coexist with relatively compact token groups when opacity is taken into account. We really appreciate the reviewer for pointing out this important distinction.


Q3: Was TokenGS trained with the same intermediate supervision? Results for TokenGS trained with this supervision, in the Table 5 setting, would largely resolve this concern.

A3: Thank you for raising this important concern. In our original experiments, TokenGS uses only final-layer supervision, whereas LocusGS receives rendering supervision at layers 6 and 12. To make a fair comparison, we additionally trained TokenGS with the same layer-6/12 intermediate supervision under the same 4-view, 256×256 base setting.

MethodSupervised LayersPSNR ↑SSIM ↑LPIPS ↓
TokenGS1221.3150.65170.4238
TokenGS + intermediate supervision6, 1222.8780.72420.3484
LocusGS6, 1224.2840.78430.2709

As shown, intermediate supervision improves TokenGS by 1.563 dB. With supervision matched, LocusGS still achieves a 1.406 dB PSNR gain. This indicates that intermediate supervision explains part, but not all, of the improvement. We appreciate the reviewer for pointing out this important factor. These additional results help disentangle the contribution of intermediate supervision from that of our proposed spatial grounding.


Q4.1: Evaluate robustness to pose noise using the TokenGS protocol.

A4.1: Thank you for this suggestion. Following the pose-noise setting described in TokenGS, we evaluate two-view RE10K reconstruction by applying random-axis rotations of 1°–10° to the non-reference input camera, while keeping the reference and target cameras unchanged. We also include a noise-free baseline.

The accompanying PSNR curves in updated paper (Figure 10) show that both methods degrade as pose noise increases. LocusGS achieves 28.60 dB with clean poses, compared with 26.33 dB for TokenGS. At 1°, the results decrease to 22.93 versus 22.58 dB, and become close at larger perturbations. These results indicate that LocusGS retains competitive reconstruction quality under pose perturbations, but its substantial advantage with accurate poses narrows considerably. We therefore do not claim improved pose-noise robustness: the larger degradation relative to its clean-pose performance highlights the dependence of our explicit geometric guidance on reliable camera parameters. We appreciate the reviewer’s suggestion and include this evaluation and limitation in the revision.


Q4.2: Evaluate cross-dataset transfer (e.g., RE10K→ACID).

A4.2: Thank you for this suggestion. We have conducted an additional RE10K→ACID cross-dataset evaluation, directly applying models trained on RE10K to a fixed subset of ACID test scenes. Both methods use the same two input views and three target views per scene, identical preprocessing at 256×256256\times256 resolution, and the same budget of 65,536 Gaussians.

MethodPSNR ↑SSIM ↑LPIPS ↓Token-level Gaussian dispersion ↓
TokenGS26.43620.79830.32055.7741
LocusGS28.05670.84300.23960.1878

These results suggest that both the reconstruction benefits and the more compact query-associated Gaussian organization transfer to ACID under the evaluated domain shift.


AuthorsOpenReview ↗

Response to Reviewer dRN8 (Part 3 / 3)

Q4.3: Report how far the anchors move from their initialization across layers?

A4.3: Thank you for this suggestion. We have added a layer-wise displacement analysis for both LocusGS and GlobalSplat augmented with our spatial grounding design. For each decoder stage ℓ\ell, we measure

D(ℓ)=1N∑i=1N∥ai(ℓ)−ai(0)∥2,D^{(\ell)}=\frac{1}{N}\sum_{i=1}^{N}\left\|\mathbf{a}_i^{(\ell)}-\mathbf{a}_i^{(0)}\right\|_2,

averaged across evaluation scenes, where ai(0)\mathbf{a}_i^{(0)} denotes the learned initialization in the trained checkpoint. Distances are measured in normalized scene coordinates.

Fig. 11 quantitatively shows that anchors move substantially from their initialization in both architectures. In LocusGS, displacement generally increases across 12 decoder layers, with minor fluctuations indicating further spatial readjustment. In GlobalSplat, displacement increases across four refinement rounds, with a smaller change in the final round.

Fig. 5 provides complementary visual evidence, showing how GlobalSplat's anchors progressively spread and reorganize toward the scene structure across the four rounds, with Pi3-predicted geometry included as a reference. These results show that the anchors are actively refined by multi-view evidence, rather than remaining near their learned initial positions.


Thank you for your time and effort in reviewing our paper, which helped us significantly strengthen the paper through these additional experiments and analyses. We hope our responses have satisfactorily addressed all of your concerns.

Thank you for the response and the additional experiment. After reading the author's reply, I have decided to maintain my original score. The doubts about the intermediate layer supervision have been largely resolved. Please incorporate the analysis of opacity weighting, the source of checkpoints, and the discussion of related work into the paper.

Reviewer_TA3pOpenReview ↗
Rating
3(+1)
Confidence
3(−1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LocusGS addresses the weak spatial coherence of Gaussian groups produced by latent queries in feed-forward 3D Gaussian Splatting. It equips each query with a learnable 3D anchor center and radius, refines this state across decoder layers, uses anchor-to-ray proximity to bias cross-attention, and predicts Gaussian centers as anchor-relative offsets. On posed RealEstate10K and DL3DV inputs, the method improves novel-view synthesis over TokenGS at matched token and Gaussian counts. The paper also reports more compact token-level Gaussian groups.

Critical strengths
  1. A focused and coherent architectural contribution. The anchor center and radius are used consistently in query interaction, cross-view aggregation, and Gaussian decoding. The component ablations support their practical value: removing the geometric cross-attention bias reduces PSNR from 24.284 to 22.751, while predicting free Gaussian centers reduces it to 22.079 (Table 5).

  2. Consistent gains against the most directly comparable baseline. With matched token and Gaussian counts, LocusGS improves DL3DV PSNR over TokenGS by 1.32–1.62 dB across the reported input-view settings, and improves RealEstate10K PSNR by 0.48 dB at both token counts (Tables 1–2). The fixed Gaussian budget is a useful property of this approach.

Critical weaknesses
  1. The comparison appears to focus on the original TokenGS, while its subsequently released latent-token version reports stronger reconstruction results. The authors should compare against this updated version under a matched evaluation protocol, separating feed-forward inference from test-time optimization. More importantly, adapting the ray-anchor mechanism to the updated latent-token architecture and measuring its incremental gain would establish whether LocusGS remains beneficial on top of the stronger model.

  2. The authors should report depth or 3D reconstruction accuracy against reference geometry and compare it with TokenGS under the same Gaussian budget. It would also be useful to test whether the anchors can be initialized or conditioned on an externally supplied point cloud or voxel representation, such as a VGGT prediction, and whether this improves performance after accounting for the additional prior and computation.

  3. The original TokenGS demonstrates dynamic scene reconstruction, whereas LocusGS reports only static-scene results. It remains unclear how ray anchors would represent geometry that changes over time. The authors should include at least one dynamic rendering experiment under a setting comparable to TokenGS, report dynamic novel-view synthesis results, and explain how the anchors evolve over time. This would test whether the proposed representation extends beyond static scenes.

Questions for authors

Please address the aforementioned weakness to support a potential increase in rating.

4 comments
AuthorsOpenReview ↗

Response to Reviewer TA3p (Part 1 / 3)

We are truly grateful for your insightful review and for noting LocusGS’s consistent gains over TokenGS across benchmarks. Your feedback has been invaluable in helping us clarify our design choices and strengthen our empirical validation. We hope the new results and explanations below fully address your questions, and we kindly ask that you consider raising your rating in light of these additions. Below, we provide detailed responses to your questions.


Q1: The comparison appears to focus on the original TokenGS, while its subsequently released latent-token version reports stronger reconstruction results. The authors should compare against this updated version under a matched evaluation protocol, separating feed-forward inference from test-time optimization. More importantly, adapting the ray-anchor mechanism to the updated latent-token architecture and measuring its incremental gain would establish whether LocusGS remains beneficial on top of the stronger model.

A1: Thank you for highlighting the updated TokenGS model. We agree that comparison with its latent-token version is important, and that feed-forward inference and test-time optimization should be reported separately.

(1) Comparison with the updated TokenGS baseline.

We evaluated the publicly released latent-token checkpoint on the same DL3DV benchmark. We report feed-forward and TTT results separately below.

Inference modeMethodParametersPSNR ↑SSIM ↑LPIPS ↓
Feed-forwardTokenGS, updated latent-token version362.7M24.460.7910.265
Feed-forwardLocusGS, original architecture241.5M24.800.8120.248
TTTTokenGS, updated latent-token version362.7M25.390.8200.238
TTTLocusGS, original architecture241.5M25.390.8280.231

For TTT, both methods perform 50 optimization steps, using only the input images as supervision and keeping network weights frozen. TokenGS updates scene-specific copies of its GS query tokens; LocusGS additionally updates the tokens with both anchor positions and radii. The appendix provides a detailed comparison of LocusGS TTT variants: optimizing query tokens alone, tokens with anchor positions, and tokens with both anchor positions and radii.

The updated TokenGS provides a stronger baseline than its original version. Nevertheless, LocusGS achieves better feed-forward reconstruction quality across all three metrics with approximately 33% fewer parameters. With TTT, the two methods achieve essentially identical PSNR, while LocusGS retains better SSIM and LPIPS.

(2) Adapting anchors to the latent-token architecture.

Following the reviewer’s suggestion, we also investigate whether anchor-based spatial grounding remains beneficial within a latent-token architecture. We retain the two-stage design of scene aggregation and Gaussian prediction, and introduce anchor positional encoding, position/radius refinement, and anchor-centered decoding in the GS decoder. Cross-attention to scene latents remains content-only because these latents do not have explicit spatial coordinates.

Full official backbone. Starting from the released checkpoint, we fine-tune the anchor-adapted model while retaining the official 12-layer encoder, 12-layer scene bottleneck, and single-layer GS decoder. Under the same feed-forward evaluation protocol, PSNR improves from 24.46 to 24.80 dB, SSIM from 0.791 to 0.802, and LPIPS from 0.265 to 0.263. These modest improvements reflect the combined effect of adaptation and additional fine-tuning.

Controlled architectural comparison. To further examine the adaptation under shared training conditions, we train two models from scratch using 3 image-encoder layers, 8 scene-bottleneck layers, and 4 GS-decoder layers. Both use the same training configuration, four input views at 256×256256\times256, and 1,024 GS queries producing 65,536 Gaussians. They differ in whether the GS decoder incorporates the anchor-based design. Both are evaluated on all 140 DL3DV test scenes without TTT.

MethodPSNR ↑SSIM ↑LPIPS ↓
TokenGS latent-token architecture20.180.5800.520
TokenGS latent-token architecture + anchor-based decoder21.720.6590.424

The anchor-based model improves PSNR by 1.54 dB, with corresponding improvements in SSIM and LPIPS. Mean token-level Gaussian dispersion also decreases from 5.5907 to 0.3376, providing complementary evidence of more compact query-associated Gaussian groups. This controlled architecture differs from the full released model; we report it separately to clarify the scope of the comparison.

AuthorsOpenReview ↗

Response to Reviewer TA3p (Part 2 / 3)

A1: (3) Additional evidence beyond TokenGS.

We additionally integrate spatial grounding into GlobalSplat (Itkin et al., 2026), incorporating geometric attention, progressive anchor refinement, and anchor-centered Gaussian decoding. Implementation details are provided in Section 3.2 of the revised manuscript.

MethodPSNR ↑SSIM ↑LPIPS ↓
GlobalSplat23.02680.68740.2751
GlobalSplat + spatial grounding26.90580.83660.1676

The query-colored visualizations in Figure 5 further illustrate more spatially concentrated Gaussian groups following the scene geometry. We also visualize the anchors at initialization and after each of the four refinement rounds to show their scene-conditioned adaptation.

We appreciate the reviewer’s suggestion. These experiments provide evidence of benefits within a latent-token design and in another query-based reconstruction method, while also showing that the gain from adapting only the final single-layer decoder of the full official model is more limited.


Q2.1: The authors should report depth or 3D reconstruction accuracy against reference geometry and compare it with TokenGS under the same Gaussian budget.

A2.1: Thank you for this suggestion. We have added a quantitative evaluation of 3D geometric accuracy, measuring how closely the predicted Gaussian centers align with reference scene geometry.

We construct reference point clouds by back-projecting the publicly released DA3-aligned depth maps from Puffin-World (Liao et al., 2026), followed by spatial subsampling. These depth-derived point clouds serve as shared reference geometry for both methods. TokenGS and LocusGS are evaluated using identical input views and the same budget of 65,536 Gaussians per scene.

Let PP denote the reference points and GG the predicted Gaussian centers. We measure the fraction of centers whose nearest reference point lies within distance δ\delta:

Aδ=1∣G∣∑g∈GI[min⁡p∈P∣g−p∣2<δ]A_{\delta}=\frac{1}{|G|}\sum_{g\in G}\mathbb{I}[\min_{p\in P}\\|g-p\\|_2 < \delta]

Higher accuracy indicates that more Gaussian centers lie close to the reference geometry. Thresholds are measured in normalized scene coordinates.

Method[email protected] ↑[email protected] ↑
TokenGS21.78%29.43%
LocusGS78.18%85.16%

Under the same Gaussian budget, LocusGS places substantially more centers near the reference geometry, providing quantitative support for improved geometric alignment beyond rendering quality alone. We further illustrate this phenomenon in Figure 12, where query-colored Gaussian centers are overlaid on reference geometry, with close-up views highlighting off-surface outliers and local Gaussian organization.


Q2.2: It would also be useful to test whether the anchors can be initialized or conditioned on an externally supplied point cloud or voxel representation, such as a VGGT prediction, and whether this improves performance after accounting for the additional prior and computation.

A2.2: Thank you for this insightful suggestion. We investigated external geometric initialization by replacing the randomly initialized anchors with points sampled from Pi3-predicted point maps using farthest point sampling (FPS).

Interestingly, the two variants exhibit nearly identical training convergence curves and final reconstruction performance. Thus, external geometric initialization provides no observable benefit in either convergence speed or reconstruction quality in our setting, despite requiring an additional geometry prediction model.

This finding is particularly interesting in comparison with UniQueR (ECCV'26), where random-only initialization reduces PSNR from 20.23 to 12.11 dB. Unlike UniQueR, which predicts a one-shot query deformation after its Transformer, LocusGS progressively refines anchor centers and adaptive support radii, allowing the updated spatial states to guide subsequent geometric attention. We hypothesize that this iterative spatial refinement reduces sensitivity to anchor initialization, enabling LocusGS to learn effective spatial representations without relying on external geometric priors. These results suggest that accurate point-cloud initialization is not necessary for LocusGS under the evaluated setting.


AuthorsOpenReview ↗

Response to Reviewer TA3p (Part 3 / 3)

Q3: It remains unclear how ray anchors would represent geometry that changes over time. The authors should include at least one dynamic rendering experiment under a setting comparable to TokenGS, report dynamic novel-view synthesis results, and explain how the anchors evolve over time. This would test whether the proposed representation extends beyond static scenes.

A3: Thank you for raising this important question. We have implemented a time-conditioned extension of LocusGS and conducted an additional dynamic reconstruction experiment against TokenGS.

Experimental setting. Since the exact training/test scene lists underlying TokenGS’s reported dynamic results were unavailable to us, we constructed a controlled comparison using a selected subset of Kubric-4D. Both methods are initialized from their respective DL3DV-trained static models and fine-tuned with four input views at 256×256256\times256 resolution, using 1,024 static queries and 256 additional dynamic queries, each producing 64 Gaussians. Evaluation uses identical camera/time samples on eight scenes: the inputs observe four cameras at different times, while the targets observe those cameras at a shared target time.

Time-dependent anchors. To extend spatial grounding to dynamic scenes, we model each dynamic anchor as a time-dependent displacement from a reference position. For query ii at decoder layer ll, we define

ail(t)=bil+Fil(t)−Fil(t0)a_i^l(t)=b_i^l+F_i^l(t)-F_i^l(t_0)

where bilb_i^l denotes the reference-time anchor position, and Fil(t)F_i^l(t) is a learned displacement function conditioned on the query feature and time embedding. The subtraction ensures that the displacement is zero at the reference time t0t_0. The support radius is conditioned on time in the same manner.

Both the reference anchor states and query features are progressively refined across decoder layers. Geometric attention uses anchor states evaluated at each input view's timestamp, while Gaussian decoding uses the states at the target timestamp. Static queries remain time-independent.

MethodPSNR ↑SSIM ↑LPIPS ↓Token-level Gaussian dispersion ↓
TokenGS24.54510.73910.36410.6195
LocusGS25.57020.81760.26810.0511

Under the same budget, LocusGS improves PSNR by 1.03 dB and reduces LPIPS by 26.4%, while maintaining substantially more compact query-associated Gaussian groups. The accompanying visual comparisons in updated paper further illustrate the reconstruction differences. These preliminary results support the feasibility of extending ray anchors to time-varying geometry. The detail are presented in updated paper Sec 2.2.8, Figure 7 and Table 4.


We truly appreciate your thoughtful comments and the opportunity to clarify our work. Please let us know if you have any remaining questions—we would be happy to discuss them further. Thank you again for your time and consideration.

Thank you for the detailed rebuttal and additional experiments. I attempted to reproduce the training progress of LocusGS. Under my setup, I observed spatial structures similar to those shown in the paper. LocusGS-base performed clearly better than TokenGS-base and converged faster, consistent with the authors’ findings. However, my reproduced LocusGS-base performed slightly below TokenGS-latent. I appreciate the authors’ new latent-architecture experiment, which reports a gain from adding anchors, but the margin over TokenGS-latent is much smaller than the original margin over TokenGS-base.

My geometric visualizations are also consistent with the paper: LocusGS produces a more organized spatial structure than TokenGS. The new quantitative analysis is helpful. I was surprised that initializing anchors from Pi3 point maps brought no clear benefit. One possible explanation is that ray-anchor attention already provides strong spatial guidance, reducing the value of this particular initialization. This is only a hypothesis, and I encourage the authors to investigate it further.

The dynamic experiment addresses my request for an initial comparison with TokenGS. The description is still brief, especially regarding the implementation of the time-dependent displacement function. Nevertheless, I accept this as an additional, preliminary result. The paper’s most thoroughly supported contribution remains in static reconstruction.

Overall, I consider this a good and reproducible piece of work. Its contribution is clear: it gives TokenGS-style queries explicit spatial grounding. The new TokenGS-latent comparisons largely address my original concern about the baseline. I still have two reservations. First, the larger TokenGS-latent model achieves rendering quality close to LocusGS without explicit geometric grounding, so the rendering gain is less significant than the original TokenGS-base comparison suggested. Second, the dynamic evaluation remains preliminary. I therefore raise my rating and adjust my confidence. I encourage the authors to release the code for the latent-token experiments.

Reviewer_P1SCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes LocusGS, a spatially grounded extension of query-based feed-forward 3D Gaussian Splatting. Each Gaussian query is augmented with a learnable 3D anchor center and support radius. The anchor state is refined across decoder layers and is used in self-attention, anchor-to-ray cross-attention, and Gaussian position decoding. Gaussian centers are predicted as radius-scaled offsets from the final anchor.

Critical strengths
  1. The problem the paper studies is reasonable. Without an explicit spatial state, Gaussians generated by the same query can be scattered across the scene.

  2. The improvements over TokenGS are consistent across RE10K and DL3DV and across PSNR, SSIM, and LPIPS.

Critical weaknesses
  1. The main concern is novelty. Scaffold-GS already organizes local Gaussians around spatial anchors. AnchorSplat also uses geometric anchors in feed-forward 3DGS. LocusGS combines anchor centers and radii with query refinement, geometric attention, and local Gaussian decoding.

  2. It is difficult to tell how much of the improvement comes simply from anchor-centered Gaussian decoding. Table 5 shows that removing free Gaussian centers causes the largest performance drop, but there is no simpler TokenGS baseline that only adds anchor-centered decoding without the full anchor-aware attention and refinement pipeline. It is hard to judge the contribution of the main spatial grounding idea.

  3. The ablations are not performed under exactly the same setting as the main DL3DV result. Table 5 uses the 256×256 base model, while the main DL3DV comparison uses the 448×256 finetuned model. It is also not clear whether TokenGS uses the same intermediate-layer rendering supervision.

  4. The paper shows improvement in within-query compactness, but the broader claim that different queries cover complementary parts of the scene is not directly measured. Compact Gaussian groups do not mean good global coverage or low redundancy between queries. An overlap or coverage analysis is required to make such a claim.

  5. The evaluation is centered on TokenGS as the controlled query-based baseline. How about other query-based 3DGS methods? Does the spatial grounding help?

Questions for authors
  1. How is LocusGS technically different from AnchorSplat and Scaffold-GS beyond using the anchor as a dynamically refined query state?

  2. How does a simpler TokenGS variant with only anchor-centered Gaussian decoding perform, without anchor-aware self-attention, anchor-to-ray cross-attention, and dynamic radius refinement?

  3. Are TokenGS and LocusGS trained with the same intermediate-layer rendering supervision in the main experiments? If not, how much of the gain comes from this difference?

  4. Can the authors provide a quantitative analysis of overlap and coverage between different query-associated regions, rather than only within-query Gaussian dispersion?

  5. Does the proposed spatial grounding improve another query-based Gaussian reconstruction method besides TokenGS?

6 comments
AuthorsOpenReview ↗

Response to Reviewer P1SC (Part 1 / 3)

We really thank the reviewer for the detailed, meaningful annd constructive feedback. We appreciate your recognition of the consistent improvements over TokenGS across different datasets and evaluation metrics.

We believe that our responses and additional experimental evidence have adequately addressed your concerns. We would sincerely appreciate it if you're confortable to reconsider your rating in light of these clarifications and new results. Below, we provide detailed responses to your questions.


Q1: How is LocusGS technically different from AnchorSplat and Scaffold-GS beyond using the anchor as a dynamically refined query state?

A1: Thank you for this important question. We clarify the differences in terms of the reconstruction paradigm and spatial query design.

(1) Comparison with AnchorSplat. The core difference lies in shifting from rigid external geometry to learning 3D spatial structure from scratch. Our design differs in three ways: First, AnchorSplat relies on pretrained point clouds with fixed anchor positions. LocusGS starts from random, learnable anchors and lets multi-view evidence pull them into coherent 3D scene layouts without external geometric priors (Figs. 1 and 5). Second, we introduce a learnable support radius that dynamically governs both geometric cross-attention and local Gaussian decoding, which proves useful in our ablations. Third, instead of front-loading features onto fixed anchors, we maintain an image–query cross-attention loop with an explicit anchor-to-ray geometric bias.

Overall, LocusGS makes discovering where to operate in 3D an organic part of feed-forward reconstruction rather than outsourcing it to external priors.

(2) Comparison with Scaffold-GS: Reconstruction Paradigm.

The major difference is that Scaffold-GS is an optimization-based method that requires thousands of iterations to fit each new scene. In contrast, LocusGS reconstructs unseen scenes in a single forward pass without any per-scene optimization, substantially reducing reconstruction time from minutes to seconds.


Q2–Q3: How does a simpler TokenGS variant with only anchor-centered Gaussian decoding perform? Are TokenGS and LocusGS trained with the same intermediate-layer supervision, and how much does this difference contribute?

A2-A3: Thanks for raising this — disentangling the impact of training supervision from our architectural design is a very fair and important check. In our initial setup, TokenGS was supervised only at the final layer (layer 12), whereas LocusGS received rendering supervision at both layers 6 and 12. To put them on completely equal footing, we ran two controlled baselines under the exact same 4-view feed-forward setting:

MethodSupervised layersPSNR ↑SSIM ↑LPIPS ↓
TokenGS1221.3150.65170.4238
TokenGS + intermediate supervision6, 1222.8780.72420.3484
Above + anchor-centered decoding6, 1223.1790.73870.3292
LocusGS6, 1224.2840.78430.2709

From these controls, two clear takeaways emerge:

  • Intermediate supervision helps, as expected — but doesn't close the gap. Aligning the supervision (layers 6 & 12) gives TokenGS a boost, so we appreciate you pointing out this difference. However, even with identical supervision, LocusGS still leads by clear gains.

  • Anchor-centered decoding is beneficial. This variant uses content-only self- and cross-attention, with no XYZ refinement or radius state/refinement; it retains the default anchor visibility regularizer. The complete LocusGS model further improves PSNR.

Overall, while intermediate supervision and anchor-centered decoding both contribute, the complete spatial interaction and refinement pipeline provides further gains. We thank the reviewer for prompting this helpful controlled analysis.


AuthorsOpenReview ↗

Response to Reviewer P1SC (Part 2 / 3)

Q4: Can the authors provide a quantitative analysis of overlap and coverage between different query-associated regions, rather than only within-query Gaussian dispersion?

A4: Thanks for this great suggestion—looking purely at within-query dispersion indeed misses how queries collaborate across the whole scene. To quantify their collective 3D organization, we added evaluations of geometric coverage, accuracy, and local inter-query mixing, along with query-colored visualizations (Figure 15).

(1) Reference geometry.

We construct reference point clouds by back-projecting the publicly released DA3-aligned depth maps from Puffin-World (Liao et al., 2026). We spatially subsample these points to reduce sampling-density bias and use the same reference geometry for both methods.

(2) Geometric coverage and accuracy.

Following the bidirectional nearest-neighbor metrics commonly used for point-cloud completeness and accuracy, we evaluate both geometric coverage and the proximity of predicted Gaussian centers to the reference geometry. Let PP denote the reference points, GG the predicted Gaussian centers, and d(x,S)d(x,S) the minimum Euclidean distance from point xx to set SS. We define coverage and accuracy as:

Cτ=1∣P∣∑p∈PI[d(p,G)<τ]C_{\tau}=\frac{1}{|P|}\sum_{p\in P}\mathbb{I}[d(p,G) < \tau]

Aδ=1∣G∣∑g∈GI[d(g,P)<δ]A_{\delta}=\frac{1}{|G|}\sum_{g\in G}\mathbb{I}[d(g,P) < \delta]

Coverage measures the fraction of reference points covered by nearby Gaussian centers, while accuracy measures the fraction of predicted centers close to the reference geometry. We classify centers within distance δ\delta as inliers and the remaining centers as outliers, giving an outlier ratio of 1−Aδ1-A_{\delta}.

Threshold τ=δ\tau=\deltaMethodCoverage ↑Accuracy ↑Outlier ratio ↓
0.05TokenGS80.05%21.78%78.22%
0.05LocusGS81.97%78.18%21.82%
0.10TokenGS91.38%29.43%70.57%
0.10LocusGS89.18%85.16%14.84%

Takeaway: While both methods cover roughly 90% of the surface at τ=0.10\tau=0.10, they achieve this very differently. Over 70% of TokenGS's Gaussians are off-surface outliers, whereas LocusGS places 85.16% of its centers directly on the geometry—showing far more disciplined budget allocation.

(3) Inter-query overlap.

To measure local spatial overlap across different query groups, we introduce Mixing@kk: the average fraction of a Gaussian’s kk nearest neighbors that belong to other queries. This provides a local measure of how strongly different query groups are spatially intermingled.

Let qiq_i be the query ID of center gig_i, and NkS(i)\mathcal{N}_k^S(i) its kk nearest neighbors within the evaluated set SS, excluding itself:

Mixing@k(S)=1∣S∣∑i:gi∈S1k∑j∈NkS(i)I[qj≠qi]\mathrm{Mixing@}k(S)=\frac{1}{|S|}\sum_{i:g_i\in S}\frac{1}{k}\sum_{j\in\mathcal{N}_k^S(i)}\mathbb{I}[q_j\neq q_i]

Lower values indicate less local cross-query mixing. We compute this measure separately within the inlier and outlier groups, using δ=0.10\delta=0.10:

RegionMethodMixing@8 ↓Mixing@16 ↓Mixing@32 ↓
InliersTokenGS68.75%74.36%81.45%
InliersLocusGS51.43%52.88%55.88%
OutliersTokenGS98.16%98.55%98.88%
OutliersLocusGS33.47%36.05%41.03%

Takeaway: LocusGS maintains much cleaner spatial boundaries between queries—especially among outliers, where TokenGS degenerates into a near-random soup of overlapping queries (98% mixing), while LocusGS preserves coherent local clusters (33–41%).

(4). Visual Confirmation. Figure 15 illustrates this difference by coloring Gaussians according to their parent query. While TokenGS leaves a large number of mixed-query Gaussians scattered away from the surfaces, LocusGS keeps its queries much closer to the scene geometry and forms more distinct local groups. The surface insets also show more complete coverage of the reference geometry by LocusGS.

We really appreciate the reviewer’s suggestion, which motivated this additional analysis. These evaluations complement within-query dispersion by examining how different queries collectively cover the geometry and organize their Gaussians in 3D space.


AuthorsOpenReview ↗

Response to Reviewer P1SC (Part 3 / 3)

Q5: Does the proposed spatial grounding improve another query-based Gaussian reconstruction method besides TokenGS?

A5: Thank you for this helpful suggestion. To examine its applicability beyond TokenGS, we additionally integrate spatial grounding into GlobalSplat (ECCV 2026), incorporating geometric attention, progressive anchor refinement, and anchor-centered Gaussian decoding(implementation details have been added to Section 3.2 of the updated paper).

MethodPSNR ↑SSIM ↑LPIPS ↓
GlobalSplat23.02680.68740.2751
GlobalSplat + spatial grounding26.90580.83660.1676

The query-colored Gaussian visualizations in Figure 5 provide complementary qualitative evidence. Each color identifies Gaussians generated by one query. Compared with the more diffuse distribution of GlobalSplat, adding spatial grounding produces a more compact distribution that more closely follows the scene geometry, with fewer scattered centers away from the apparent surfaces.

We also visualize the anchor positions at initialization and after each of the four refinement rounds in Figure 5. This reveals how the shared initial anchors progressively adapt to the input scene through interaction with image evidence. Together, these results support the applicability of our spatial grounding design to another query-based Gaussian reconstruction architecture.

Itkin, Roni, et al. "Globalsplat: Efficient feed-forward 3d gaussian splatting via global scene tokens." European Conference on Computer Vision. Cham: Springer Nature Switzerland, 2026.


Thank you again for your time, effort, and constructive suggestions. We hope that the new experimental evidence and clarifications above have resolved your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.