acceptodds
Under review as a conference paper at ICLR 2027

Per-View Gaussian Predictions Enable Training-Free Distractor Filtering in Feed-Forward 3DGS

Abstract

Feed-forward 3D Gaussian Splatting reconstructs an explicit Gaussian representation from multiple input images, making 3D reconstruction increasingly accessible for casual captures. However, such captures frequently contain transient content that is inconsistent across the input views. Such content can persist in the reconstructed Gaussian representation despite conflicting with observations from other views. As a result, it may produce blurred, duplicated, or floating artifacts in novel views. We introduce a training-free filtering procedure that exploits the association between predicted Gaussians and their corresponding input views. Self-view filtering compares each input with a rendering that excludes its associated Gaussians, exposing content unsupported by the remaining Gaussians. Cross-view filtering measures how those Gaussians interfere with reconstruction of other input views. Gaussians selected by either filtering path are removed together. The procedure operates on a single frozen prediction without retraining or scene-specific optimization. Evaluations across recent feed-forward 3DGS models on novel view synthesis benchmarks with distractors in the input images show improved rendering quality with varying numbers of input views. On clean inputs, the original reconstructions are largely preserved.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Computer Vision, Distractor-Free Reconstruction, Feedforward 3DGS

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_2X6WOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a training-free distractor filtering method for a feedforward 3DGS model, which utilizes per-view Gaussian exclusion and cross-view distractor to identify the Gaussian components to be removed. Experiments conducted on various baseline models and two distractor benchmark datasets demonstrate that this method significantly improves the performance of novel views while causing only a limited decline in performance for clean inputs.

Critical strengths

This method provides a practical filtering process that does not require retraining or optimization for specific scenarios. Comparisons across multiple backbone networks and varying numbers of inputs show that this method achieves a consistent improvement in average performance across all datasets.

Critical weaknesses
  1. Compared with simple differential-based filtering methods, the advantages of this approach have not yet been fully validated. Currently, there is a lack of benchmark methods based on fully reconstructed residuals, and as shown in Table 6, adding self-view filtering to cross-view filtering only improves the average PSNR by 0.15–0.18 dB.

  2. Evidence regarding robustness is rather limited. The evaluation of noise-free scenes covered only four scenes, and the method’s sensitivity to thresholds and depth/opacity protection measures was not investigated.

Questions for authors
  1. How does the method compare with full-reconstruction discrepancy filtering using the same features, validity conditions, and a comparable pruning budget?

  2. How were the thresholds selected, and can threshold ablations clarify whether the reported gains and clean-scene preservation remain stable?

10 comments

W1 & Q1.

(W1) Compared with simple differential-based filtering methods, the advantages of this approach have not yet been fully validated. Currently, there is a lack of benchmark methods based on fully reconstructed residuals, and as shown in Table 6, adding self-view filtering to cross-view filtering only improves the average PSNR by 0.15–0.18 dB.

(Q1) How does the method compare with full-reconstruction discrepancy filtering using the same features, validity conditions, and a comparable pruning budget?

Thank you for this constructive suggestion. The paper states that the discrepancy between renderings and inputs can be obscured in the feed-forward setting, but we did not support this with an experiment.

We tested the baseline you describe, which compares each input with the rendering of all Gaussians, as optimization-based distractor-free methods commonly do. It uses the same features, opacity condition and pruning budget as our method. We also report its best single threshold, and our self-view path alone under the same budget. We have added this comparison to the paper. The table gives the change in PSNR averaged over the three models from our main experiments.

4 inputs8 inputs16 inputs
residual on the full rendering, same budget−0.06+0.14+0.19
residual on the full rendering, best threshold+0.03+0.08+0.18
our self-view path alone, same budget+0.75+0.72+0.63
ours+0.96+1.11+0.98

Regarding the gain from adding self-view to cross-view filtering, both paths build on the same idea of using the per-view Gaussian association to expose inconsistency across views. They examine a view’s Gaussians from that view’s camera and from the other input cameras, respectively, so their selections naturally overlap in part. Combining the two signals consistently gives the best results.

W2 & Q2.

(W2) Evidence regarding robustness is rather limited. The evaluation of noise-free scenes covered only four scenes, and the method’s sensitivity to thresholds and depth/opacity protection measures was not investigated.

(Q2) How were the thresholds selected, and can threshold ablations clarify whether the reported gains and clean-scene preservation remain stable?

We use the same thresholds, τ_self = 0.5 and τ_cross = 0.05, for all models, datasets and input counts. They were chosen empirically from a few candidate values and then fixed. The table gives the change in PSNR on cluttered / clean inputs over a range of values with 4 inputs, averaged over the three models from our main experiments. Cluttered inputs cover the scenes of RobustNeRF and NeRF On-the-go, and clean inputs the four RobustNeRF scenes.

τ_self \ τ_cross0.010.0250.05 (ours)0.10.2
0.3+0.51 / −0.45+0.62 / −0.35+0.66 / −0.32+0.66 / −0.30+0.62 / −0.30
0.4+0.77 / −0.24+0.87 / −0.14+0.90 / −0.11+0.88 / −0.08+0.81 / −0.08
0.5 (ours)+0.85 / −0.16+0.94 / −0.07+0.96 / −0.04+0.91 / −0.02+0.76 / −0.02
0.6+0.79 / −0.14+0.87 / −0.05+0.86 / −0.02+0.74 / 0.00+0.44 / 0.00
0.7+0.76 / −0.14+0.83 / −0.05+0.81 / −0.02+0.68 / 0.00+0.34 / 0.00

For the opacity condition, we changed its threshold and also removed it. The table gives the change in PSNR on cluttered / clean inputs in the same setting.

self-view opacity thresholdremoved0.10.30.5 (ours)0.70.9
ΔPSNR (dB)+0.97 / −0.04+0.97 / −0.04+0.97 / −0.04+0.96 / −0.04+0.95 / −0.03+0.92 / −0.02

For the depth conditions, we added a relative margin of ±5% and ±10% of the depth and also removed each condition.

depth marginremoved−10%−5%0 (ours)+5%+10%
self-view depth+0.80 / −0.05+0.92 / −0.04+0.93 / −0.04+0.96 / −0.04+0.97 / −0.04+0.97 / −0.04
cross-view depth−0.19 / −0.08+0.89 / −0.06+0.93 / −0.05+0.96 / −0.04+0.95 / −0.02+0.90 / −0.02

For self-view filtering, we also removed the opacity and depth conditions together.

self-view conditions removednone (ours)opacitydepthboth
ΔPSNR (dB)+0.96 / −0.04+0.97 / −0.04+0.80 / −0.05+0.07 / −0.80

Around our settings, the gains are stable and clean inputs remain close to the unfiltered results. The conditions are needed as a whole. Removing the cross-view depth condition, or both self-view conditions, loses most of the gain.

We also evaluated 12 more distractor-free scenes from Mip-NeRF 360 and Tanks and Temples. The table gives the change in PSNR averaged over the same three models.

4 inputs8 inputs16 inputs
RobustNeRF, 4 scenes−0.04−0.01+0.06
Mip-NeRF 360, 6 scenes−0.07−0.01+0.06
Tanks and Temples, 6 scenes0.00−0.01+0.11

Quality is largely preserved on all three datasets. We have added these results to the paper.

Thank you for the additional experiments. The robustness results address my second concern. Could you clarify which depth validity checks were applied to the full-rendering baseline?

Reviewer_sd7mOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a training-free method for removing distractor Gaussians from feed-forward 3D Gaussian Splatting reconstructions. The key insight is that existing models natively associate each predicted Gaussian with a specific input view. The authors repurpose this association as a cue for filtering: for each view's Gaussian subset, they compute (1) a self-view discrepancy score and (2) a cross-view interference score. Gaussians exceeding thresholds on either score (with opacity/depth validity conditions) are removed from the frozen prediction, and novel views are rendered directly from the pruned set. The method uses DINOv3 patch features for similarity, requires no retraining, no masks, no held-out images, and no scene optimization. Experiments span RobustNeRF and NeRF On-the-go with 4/8/16 views across four reconstruction models, showing consistent PSNR/SSIM/LPIPS gains over unfiltered baselines, plus ablation (self vs. cross vs. both, CLIP/L1 vs. DINOv3), comparison to GenWildSplat, preservation on clean scenes, and runtime overhead (0.1–2.1 s). The contribution is a simple, model-agnostic post-hoc filter exploiting an architectural property of feed-forward Gaussian predictors.

Critical strengths

(1) The core idea is elegant, well-motivated, and training-free. The observation that per-view Gaussian provenance can be inverted into a filtering signal is a clean contribution. The two complementary signals are theoretically sensible and the ablation confirms each contributes independently. This matters because it makes the method immediately applicable to any frozen feed-forward 3DGS model without retraining, masks, or optimization.

(2) The paper evaluates on two distractor benchmarks at three input counts, three-plus backbones, an in-the-wild model (AnySplat), a comparison to a trained distractor-aware method (GenWildSplat), per-scene breakdowns, clean-scene preservation tests with both GT and estimated poses, and runtime overhead. The consistent improvement and the small degradation on clean scenes strongly support the claim that gains come from filtering transient content.

Critical weaknesses

(1) The failure modes are acknowledged only briefly: the method relies on inconsistency, not semantics, so consistently-observed distractors (e.g., a person present in every view) are retained, and the paper offers no quantitative characterization of when filtering helps vs. hurts. The clean-scene preservation numbers show the filter is not free; a per-scene analysis of false removals (which static content gets pruned and why) is absent.

(2) The dependence on model-specific architecture and the narrow distractor definition limit generality. The method requires the reconstruction model to expose a per-view Gaussian association; the paper explicitly concedes this is unavailable in architectures without it, making the approach non-universal.

Questions for authors

(1) It is suggested to report, on at least one benchmark, the fraction of distractor Gaussians retained (e.g., via available masks) and the fraction of static Gaussians erroneously removed?

(2) The authors note people observed consistently across views may be retained. Can you construct or identify a setting (e.g., a static-looking but non-scene object present in all inputs) and report behavior? Does the cross-view signal ever fire for consistently-observed content, and under what geometric conditions?

(3) How were τ_self and τ_cross chosen, and how sensitive are results to them? Are the same thresholds used across all models/datasets/input counts, and does performance degrade gracefully?

(4) GenWildSplat uses AnySplat-estimated cameras. Would applying your filter to AnySplat with GT poses (where available) change the comparison, and can you clarify whether GenWildSplat's masks are being evaluated on inputs it was trained for?

(5) For architectures lacking native per-view Gaussian provenance (e.g., voxel- or token-aligned predictors), could provenance be approximated (e.g., by nearest-input-view assignment), and would the method still work? A brief experiment or discussion would strengthen the generality claim.

4 comments

W1 & W2 (definition)

(W1) The failure modes are acknowledged only briefly: the method relies on inconsistency, not semantics, so consistently-observed distractors (e.g., a person present in every view) are retained, and the paper offers no quantitative characterization of when filtering helps vs. hurts. The clean-scene preservation numbers show the filter is not free; a per-scene analysis of false removals (which static content gets pruned and why) is absent.

(W2) … the narrow distractor definition limit generality. …

Our distractor definition follows RobustNeRF and NeRF On-the-go, which characterize distractors as non-persistent or dynamic content rather than by category. It therefore covers arbitrary objects, such as the toys in RobustNeRF. Content that stays consistent across the inputs, such as a person standing still, is where a semantic prior would help instead, which makes the two cues potentially complementary.

Regarding when filtering helps, the gain follows the amount of inconsistent content. The Crab and Yoda scenes in the RobustNeRF dataset provide paired clean and cluttered photos from the same camera poses, so we replaced 0 to 4 of the 4 clean inputs with their cluttered counterparts. The table gives the change in PSNR averaged over the three models from our main experiments.

inputs with distractors0/41/42/43/44/4
ΔPSNR (dB)+0.01+0.40+0.79+1.15+1.49

Regarding when it hurts, static content can be removed when it appears in only one input, especially close to that camera, and occludes what the other inputs observe. No other view supports such content, so from the given inputs alone it is indistinguishable from a distractor captured by a single view. Resolving such cases requires information beyond the given inputs, such as additional views or a semantic prior.

We also extended the evaluation on distractor-free inputs from 4 to 16 scenes. The table gives the change in PSNR averaged over the same three models.

4 inputs8 inputs16 inputs
RobustNeRF, 4 scenes−0.04−0.01+0.06
Mip-NeRF 360, 6 scenes−0.07−0.01+0.06
Tanks and Temples, 6 scenes0.00−0.01+0.11

The drop on distractor-free inputs is much smaller than the gain when distractors are present. We have updated the limitations and added visualizations of the removed Gaussians on distractor-free inputs, where the removed region is observed by only one input.

Our answers to Q1 and Q2 below report what the filter removes and how it handles consistently observed content.


Q1.

It is suggested to report, on at least one benchmark, the fraction of distractor Gaussians retained (e.g., via available masks) and the fraction of static Gaussians erroneously removed?

To our knowledge, these benchmarks do not provide distractor masks, so an exact count of the distractor Gaussians that remain is not available. We instead report a measurement that needs no masks, using the experiment in W1 above, where clean inputs of Crab and Yoda are replaced by their cluttered counterparts. A clean input contains only static content, so any Gaussian removed from it is an erroneous removal. The table gives the fraction of Gaussians removed from each type of input, averaged over the same three models.

inputs with distractors0/41/42/43/44/4
removed from cluttered inputs (%)–3.743.904.074.21
removed from clean inputs (%)0.110.250.390.52–

Removal concentrates on the cluttered inputs, and few Gaussians are removed from the clean ones. We have added these results to the paper.

Q2.

The authors note people observed consistently across views may be retained. Can you construct or identify a setting (e.g., a static-looking but non-scene object present in all inputs) and report behavior? Does the cross-view signal ever fire for consistently-observed content, and under what geometric conditions?

We identified two cases in real captures in which a person, or a depiction of one, appears in all 4 inputs. One is a scene from the Deep Blending dataset with a portrait on a wall, and the other is a scene from NeRF On-the-go in which two people stand and talk. The table gives the fraction of all Gaussians removed by each path, averaged over the same three models.

removed (%)self-viewcross-view
Deep Blending0.40.0
NeRF On-the-go1.33.4

In the visualizations we have added to the paper, the portrait, which is completely static, is not filtered, while the standing people change their poses between the inputs and the parts that change are filtered. The few Gaussians removed in the first scene lie on a region that the other inputs do not observe, which is the case described in W1 above.

Regarding the cross-view signal, it measures whether a view’s Gaussians interfere with reproducing another input when they are added in that view, and it is evaluated only for Gaussians that lie in front of that view’s surface. Content observed consistently agrees with what the other inputs show, so the signal stays silent, as with the portrait. It can still fire where such content differs between the inputs, as with the standing people.


W2 (architecture) & Q5

(W2) The dependence on model-specific architecture and the narrow distractor definition limit generality. The method requires the reconstruction model to expose a per-view Gaussian association; the paper explicitly concedes this is unavailable in architectures without it, making the approach non-universal.

(Q5) For architectures lacking native per-view Gaussian provenance (e.g., voxel- or token-aligned predictors), could provenance be approximated (e.g., by nearest-input-view assignment), and would the method still work? A brief experiment or discussion would strengthen the generality claim.

We had limited the scope to models with a native per-view association. Following your suggestion, we tested whether the association can be approximated in voxel- and token-based feed-forward 3DGS models, with the same thresholds.

We compared three ways of assigning each Gaussian to an input view. The first is the nearest view, as you suggested, and the second is the view whose pixel color it matches best. The third follows the model. A voxel in VolSplat or the voxel-merged AnySplat takes the view that contributes most to it, and a ZipSplat scene token takes the view of its source patch.

The table gives the change in PSNR on cluttered / clean inputs with 4 inputs. Cluttered inputs cover the scenes of RobustNeRF and NeRF On-the-go, and clean inputs the four RobustNeRF scenes.

view assignmentVolSplatAnySplat (voxel-merged)ZipSplat
nearest view+0.10 / +0.08−0.01 / −0.09+0.08 / −0.10
color-based+0.41 / +0.21+0.14 / 0.00+0.21 / 0.00
model-based+0.86 / +0.17+0.36 / −0.02+0.69 / 0.00

The gain increases from the nearest-view to the color-based assignment, and further with the model-based one. The method can thus be applied to other architectures, depending on how their Gaussians are assigned to input views. We have added this experiment to the paper. Thank you for pointing us in this direction.

Q3.

How were τ_self and τ_cross chosen, and how sensitive are results to them? Are the same thresholds used across all models/datasets/input counts, and does performance degrade gracefully?

Yes, the same values, τ_self = 0.5 and τ_cross = 0.05, are used for all models, datasets and input counts. They were chosen empirically from a few candidate values and then fixed. Performance degrades gracefully around them. The table gives the change in PSNR on cluttered / clean inputs over a range of values with 4 inputs, averaged over the three models from our main experiments.

τ_self \ τ_cross0.010.0250.05 (ours)0.10.2
0.3+0.51 / −0.45+0.62 / −0.35+0.66 / −0.32+0.66 / −0.30+0.62 / −0.30
0.4+0.77 / −0.24+0.87 / −0.14+0.90 / −0.11+0.88 / −0.08+0.81 / −0.08
0.5 (ours)+0.85 / −0.16+0.94 / −0.07+0.96 / −0.04+0.91 / −0.02+0.76 / −0.02
0.6+0.79 / −0.14+0.87 / −0.05+0.86 / −0.02+0.74 / 0.00+0.44 / 0.00
0.7+0.76 / −0.14+0.83 / −0.05+0.81 / −0.02+0.68 / 0.00+0.34 / 0.00

The trend is similar with 8 and 16 inputs, and we have added these results to the paper.


Q4.

GenWildSplat uses AnySplat-estimated cameras. Would applying your filter to AnySplat with GT poses (where available) change the comparison, and can you clarify whether GenWildSplat's masks are being evaluated on inputs it was trained for?

We designed this comparison with fairness in mind and should have stated it more clearly. GenWildSplat extends AnySplat’s architecture and is initialized from its pretrained weights, so we apply our filter to AnySplat and the two share the same backbone.

AnySplat predicts its cameras from unposed images and cannot take known poses as input, so this comparison cannot be run with ground-truth poses. Our filter, however, is applied to a frozen AnySplat prediction, so its gain over AnySplat is measured with the same Gaussians and cameras and does not come from the poses. The difference from GenWildSplat also appears where no distractors are involved. On the distractor-free inputs of Figure 4 and Table 4 in the paper, its masks cover static objects, and its PSNR is already below the unfiltered AnySplat. With known poses, our filter also consistently improves DepthSplat, YoNoSplat and ReSplat in the main experiments.

Regarding the masks, neither model is trained on RobustNeRF or NeRF On-the-go. GenWildSplat’s masks select predefined object categories and behave as designed on these inputs. These benchmarks, however, characterize distractors as content that is not persistent or is dynamic, not by category, so the masked regions do not always coincide with them.

Thank you for the detailed response and additional experiments. I have carefully reviewed all the comments, which indeed resolve my concerns. I have no further comments and will raise my score, thanks for the efforts.

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

This work investigates an interesting problem: how to achieve feed-forward 3D Gaussian Splatting while filtering out distractors. To address this challenge, the authors propose Self-View Filtering and Cross-View Filtering to identify distractors. Experimental results demonstrate the effectiveness of the proposed method.

Critical strengths
  1. It is important to investigate the impact of distractors in feed-forward 3DGS. The authors propose a training-free paradigm to address this problem, which is an elegant approach.

  2. By combining Self-View Filtering and Cross-View Filtering, the proposed method enables feed-forward distractor filtering.

  3. Both qualitative and quantitative experiments validate the effectiveness of the proposed method.

Critical weaknesses
  1. Concerns About View Selection: My main concern is that the Self-View Filtering and Cross-View Filtering strategies depend on the selection of views for removal or addition. If the selected views contain no distractors, while distractors appear in the unselected views, can the proposed method still maintain its effectiveness?

  2. Limited Experimental Details: How many distractors are present in the different experimental settings? This information is crucial for evaluating the effectiveness of the proposed method.

Questions for authors

Robustness to Frequent Distractors: What happens when distractors appear very frequently? Can the proposed method still maintain stable performance? An ablation study may be necessary to investigate this issue.

3 comments

W1.

Concerns About View Selection: My main concern is that the Self-View Filtering and Cross-View Filtering strategies depend on the selection of views for removal or addition. If the selected views contain no distractors, while distractors appear in the unselected views, can the proposed method still maintain its effectiveness?

We apologize that this was unclear. Our method does not select views. Self-View Filtering excludes each input view's Gaussians in turn, and Cross-View Filtering evaluates each view's Gaussians against every other input view and aggregates the result over all of them. We have clarified this in the paper.

Beyond this clarification, your comment motivated us to test whether the method remains effective when distractors appear in only some of the input views. Crab and Yoda scenes in the RobustNeRF dataset provide paired clean and cluttered photos from the same camera poses, which lets us isolate the effect of distractors. We replaced 0 to 4 of the 4 clean inputs with their cluttered counterparts, in every possible combination, keeping the cameras and test views fixed. We have added this experiment to the paper. The table gives the change in PSNR averaged over the three models from our main experiments.

inputs with distractors0/41/42/43/44/4
ΔPSNR (dB)+0.01+0.40+0.79+1.15+1.49

The gain grows with the number of inputs that contain distractors.


W2.

Limited Experimental Details: How many distractors are present in the different experimental settings? This information is crucial for evaluating the effectiveness of the proposed method.

In the RobustNeRF dataset, all inputs are cluttered photos: one balloon in Statue, three wooden robots in Android, and a pool of 100+ objects in Crab and Yoda. In the NeRF On-the-go dataset, distractors include people and cars. Its six benchmark scenes consist of two each at low, medium and high occlusion levels, and the dataset reports occlusion ratios from 5% to 30%. We have updated the experimental setup in the paper accordingly.


Q1.

Robustness to Frequent Distractors: What happens when distractors appear very frequently? Can the proposed method still maintain stable performance? An ablation study may be necessary to investigate this issue.

Our main RobustNeRF experiments are already this setting: every input contains distractors, and filtering improves PSNR for all three models at 4, 8 and 16 inputs, by +1.29, +0.97 and +0.76 dB on average. The method does not need clean views, since distractors differ from view to view and thus remain inconsistent with the other views. The experiment in W1 confirms this: the gain increases monotonically with the distractor frequency.

Reviewer_dZhwOpenReview ↗

I thank the authors for their response.

I have carefully reviewed the authors' response as well as the comments from the other reviewers. The authors' response has addressed my concerns. However, I also understand the concerns raised by other reviewers regarding the priority of validating the claims made in this manuscript. Therefore, I have decided to maintain my original score and look forward to further discussions with the other reviewers and the AC.

Thank you very much for carefully reviewing our response. We are glad that it addressed your concerns, and we would be happy to provide any further clarification during the discussion.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We thank the reviewers for their careful and constructive feedback.

We are encouraged that they found the problem important (dZhw), the idea elegant (dZhw, sd7m) and well motivated (sd7m), and the method practical (2X6W) with consistent improvements (sd7m, 2X6W). The concerns centered on validation, and we have addressed them with additional experiments and clarifications. Detailed responses are posted under each review.

For convenience during the discussion period, we have placed the new experiments in Appendix C, at the end of the paper, and marked the changes to the main paper in blue. We will integrate these experiments into the corresponding sections in the camera-ready version. The revised paper contains the following.

  • Sec. 3.2 clarifies that both filtering steps use every input view (dZhw).
  • Sec. 4 describes the distractors of each benchmark (dZhw).
  • Sec. 5 extends the limitations (sd7m).
  • Appendix C.1 varies the number of inputs that contain distractors and reports the Gaussians removed from cluttered and clean inputs (dZhw, sd7m).
  • Appendix C.2 compares with filtering by the full-rendering residual (2X6W).
  • Appendix C.3 analyzes the sensitivity to the thresholds and to the opacity and depth conditions (2X6W, sd7m).
  • Appendix C.4 evaluates 12 more distractor-free scenes (2X6W, sd7m).
  • Appendix C.5 visualizes the removed Gaussians on clean inputs and on consistently observed content (sd7m).
  • Appendix C.6 applies the method to voxel- and token-based models (sd7m).

Thank you again for your time and effort.

Best regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.