acceptodds
Under review as a conference paper at ICLR 2027

Where Fixes Really Help: Cross-Fitted Repair-Model Supervision for 3DGS

Abstract

Sparse-view 3D Gaussian Splatting (3DGS) is severely underconstrained away from observed cameras, often producing unstable geometry and appearance. Pretrained repair models can enrich supervision by correcting rendered novel views, yet their outputs may also introduce scene-inconsistent content that 3DGS absorbs into persistent artifacts. The resulting challenge is to determine which repaired pixels provide reliable supervision. We introduce FixTrust-GS, a framework that estimates pixel-wise repair reliability from the captured scene. FixTrust-GS constructs cross-fitted auxiliary reconstructions using only captured images and evaluates whether each repaired pixel agrees with their predictions better than the corresponding initial render. This relative agreement improvement yields dense weights for continued 3DGS optimization, retaining supported repair content while suppressing inconsistent edits. The framework requires only rendered–repaired image pairs and works with pretrained repair models without modifying their parameters. Experiments across three datasets, two distinct repair models, and three sparse-view budgets show that FixTrust-GS improves reconstruction fidelity over uniform whole-image weighting while often using fewer Gaussians in the final model. Code: https://anonymous.4open.science/r/FixTrust-GS-1D4B.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Sparse-View Reconstruction, Image Repair, Reliability-Weighted Supervision

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_UAubOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes FixTrust-GS, a method for weighting repaired-view supervision in sparse-view 3D Gaussian Splatting. It trains auxiliary reconstructions on complementary subsets of captured images and checks whether repaired pixels agree with these reconstructions better than the original renders. The resulting pixel-wise weights are used during continued 3DGS optimization. Evaluation covers three datasets, two pretrained repair models, and three training-view budgets.

Critical strengths
  1. The paper addresses a relevant problem. Repaired images can introduce scene-inconsistent content, so deciding which edits to trust is important. Comparing a repair against the original render, rather than judging its absolute disagreement alone, is a reasonable approach. The method also works without modifying the pretrained repair model.

  2. The reliability diagnostics provide useful supporting evidence. Figure 4 examines whether higher weights correspond to actual repair improvements against ground truth. This is more informative than reporting reconstruction metrics alone and suggests that the weighting captures some useful signal.

Critical weaknesses
  1. The added value of the proposed weighting is not convincingly established.

The experiments mainly compare Train-only and different supervision weights within the same repair pipeline. These are useful controls, but they do not establish an advantage over the existing reliability-aware repair methods discussed in Section 2.

Several improvements become quite small against stronger scalar settings in the appendix. For example, Table 17 reports 17.34 versus 17.29 dB on Deep Blending with nine training views and 48 repair views, compared with a global weight of 0.1. Meanwhile, LPIPS worsens from 0.571 to 0.576, and the proposed method uses slightly more Gaussians. On LLFF with nine training views, both repair backends remain below Train-only in PSNR. These results do not demonstrate a consistent quality advantage.

The extensive parameter sweeps also do not replace a robustness evaluation: only one training seed is reported. Furthermore, fixed-HW weighting changes both the spatial allocation and the overall strength of repair supervision. Figure 4’s equal-mass diagnostic is useful, but it is a post-hoc analysis rather than an end-to-end reconstruction control. Comparisons against per-image mean weights or spatially shuffled maps would help establish whether precise pixel selection is responsible for the gains.

  1. The qualitative presentation does not adequately substantiate the artifact-reduction claims.

Figure 4 and the appendix comparisons pack many small images into dense grids. Labels and numerical overlays compete with the image content, while enlarged crops and clearly identified artifact regions are missing. At the presented scale, it is difficult to determine which structures or textures are improved relative to the baselines. Adding more thumbnails does not resolve this problem. For a paper about rejecting harmful repairs, the qualitative evidence should make the corrected and preserved details directly inspectable; the current figures do not do this convincingly.

Questions for authors
  1. Do the final reconstruction gains remain when comparing against per-image mean weights and spatially shuffled maps with the same total weight? This would help separate spatial reliability estimation from simply weakening repair supervision.

  2. How stable are the reported margins across random seeds and captured-view selections? Please also clarify how the default settings were selected and whether a separate validation set was used.

  3. How does the method compare with an existing reliability-aware repair integration method under matched conditions? Such a comparison would materially affect my assessment of the contribution.

3 comments
AuthorsOpenReview ↗

Comment 1 — Spatial weighting and qualitative evidence

We thank the reviewer for the detailed assessment of spatial weighting, robustness, and qualitative evidence. The requested controls help distinguish pixel selection from reduced supervision strength, and we report them alongside the presentation revisions.

In result tables, bold marks the best reported value per metric within each comparison, including ties.

W1. Added value and robustness

We agree that the original results do not establish a consistent advantage across all settings. In the cited Deep Blending setting, the PSNR gain over Global 0.1 is only 0.05 dB, while LPIPS and Gaussian count are worse. Both repair models also remain below Train-only in PSNR on LLFF with nine training views. We will state these trade-offs explicitly and qualify the abstract, main-results discussion, and conclusion rather than imply universal improvements.

We have completed the requested comparisons on Mip-NeRF 360, Train 9, Repair 48. Against per-image mean and shuffled weights, the PSNR gains are 0.260/0.259 dB for Difix3D and 0.088/0.094 dB for Diffusion Harmonizer, with improved mean SSIM and LPIPS (Question 1, Table 1). We also report three optimization seeds and alternative captured-view subsets (Question 2), and a matched HAD comparison with gains of 0.239/0.187 dB (Question 3). We will add these results and their evaluation scope to the paper.

W2. Inspectable qualitative evidence

The dense thumbnail grids did not make the relevant details sufficiently inspectable. We have revised Figure 4 and Appendix C.4 to show fewer, larger examples, with magnified crops beside the corresponding full images. Colored boxes mark identical regions across methods, and captions identify the structures and textures to compare. Method labels are placed outside the image content; reported PSNR is computed on the full image, not the selected crop. The comparisons retain Train-only, scalar-weight baselines, and our method. These changes make the claimed differences directly inspectable rather than relying on additional thumbnails.

Q1. Equal-total-weight reconstruction controls

Yes, the average reconstruction gains remain against both requested controls. For each repaired image, we replace its map either by its own mean weight or by a spatial permutation of its weights. Both preserve its total weight; shuffling also preserves its histogram. All methods resume the same checkpoint with the same repairs, cameras, fixed-HW denominator, and optimization schedule. These are end-to-end reconstruction runs, distinct from Figure 4's post-hoc diagnostic. Table 1 supports a benefit from spatial allocation beyond simply reducing total supervision weight. Our Gaussian counts are slightly higher than these controls, so this comparison supports improved reconstruction quality, not an additional reduction in model size.

Table 1. End-to-end equal-weight controls on Mip-NeRF 360, Train 9, Repair 48, K=3K=3. Each control preserves the total weight within each repaired image.

Difix3D

SupervisionPSNR ↑SSIM ↑LPIPS ↓Gaussians (10³)
Per-image mean17.7980.50080.4075537.6
Spatial shuffle17.7990.50070.4072538.5
FixTrust-GS18.0580.50660.4020544.9

Diffusion Harmonizer

SupervisionPSNR ↑SSIM ↑LPIPS ↓Gaussians (10³)
Per-image mean18.1070.51060.4203536.2
Spatial shuffle18.1020.51040.4199537.2
FixTrust-GS18.1950.51270.4179542.9
AuthorsOpenReview ↗

Comment 2 — Random seeds and default settings

Q2. Random seeds, captured views, and default settings

Optimization seeds. We repeated the full pipeline with seeds 42/43/44 at Train 9, Repair 48, K=3K=3, on Bicycle, Bonsai, Counter, Garden, Kitchen, Room, and Stump. Training/test identities remain fixed; the main and auxiliary models are retrained and repaired supervision is regenerated for each seed. Table 2 reports the mean and sample standard deviation of per-seed scene averages, including the full scalar sweep. Our mean PSNR, SSIM, and LPIPS improve over every tested scalar weight for both repair models. This supports robustness in the tested setting, not statistical significance for every original dataset or view budget.

Table 2. Mean ±\pm sample standard deviation over seeds 42/43/44, Train 9, Repair 48, K=3K=3, on the stated Mip-NeRF 360 subset.

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
Train-only18.335 ± 0.0810.5291 ± 0.00150.3306 ± 0.002118.325 ± 0.1110.5286 ± 0.00200.3310 ± 0.0021
Global 1.018.687 ± 0.0540.5313 ± 0.00020.3478 ± 0.001419.108 ± 0.0410.5414 ± 0.00050.3982 ± 0.0027
Global 0.918.710 ± 0.0310.5329 ± 0.00060.3477 ± 0.001019.111 ± 0.0280.5424 ± 0.00030.3966 ± 0.0018
Global 0.818.742 ± 0.0400.5346 ± 0.00030.3477 ± 0.001019.128 ± 0.0240.5435 ± 0.00030.3954 ± 0.0014
Global 0.718.755 ± 0.0470.5364 ± 0.00010.3480 ± 0.001119.140 ± 0.0260.5447 ± 0.00060.3941 ± 0.0019
Global 0.618.781 ± 0.0360.5381 ± 0.00040.3483 ± 0.000919.163 ± 0.0250.5466 ± 0.00040.3916 ± 0.0019
Global 0.518.807 ± 0.0430.5403 ± 0.00030.3488 ± 0.000819.182 ± 0.0340.5484 ± 0.00040.3890 ± 0.0024
Global 0.418.832 ± 0.0340.5428 ± 0.00050.3487 ± 0.001019.204 ± 0.0320.5508 ± 0.00050.3850 ± 0.0018
Global 0.318.855 ± 0.0580.5453 ± 0.00060.3494 ± 0.000719.224 ± 0.0370.5537 ± 0.00020.3802 ± 0.0013
Global 0.218.902 ± 0.0460.5488 ± 0.00030.3487 ± 0.001219.229 ± 0.0120.5569 ± 0.00080.3737 ± 0.0019
Global 0.118.906 ± 0.0410.5510 ± 0.00070.3496 ± 0.000619.214 ± 0.0190.5598 ± 0.00060.3645 ± 0.0018
FixTrust-GS19.163 ± 0.0660.5573 ± 0.00100.3430 ± 0.001519.317 ± 0.0320.5631 ± 0.00030.3596 ± 0.0012

Default settings and validation. The reported experiments use K=3K=3, fixed-HW normalization, and 48 repair views as common defaults. We selected K=3K=3 based on the completed fold-count ablation and its quality–cost trade-off. Fixed-HW retains the full-image denominator so that reliable area affects a repair's total contribution. The supplementary seed and view-selection comparisons hold these settings fixed. The available experiment records define training, repair, and test splits, but do not document a separate validation set. We therefore do not present these defaults as independently validation-selected. We will explicitly document the selection procedure and distinguish the reported ablations from independent validation.

AuthorsOpenReview ↗

Comment 3 — Captured-view selection and HAD comparison

Q2. Captured-view selection (continued)

Captured-view selection. We evaluated the original subset and two alternatives selected by random-start, pose-based farthest-point sampling (selection seeds 137/271), with optimization seed 42 fixed. The test images, non-test candidate pool, Train 9 and Repair 48 counts remain fixed. The remaining non-test cameras become repair views, so repair-camera identities can change with the training subset. No image-quality or test metric selects the subsets. All models and repairs are regenerated for each alternative, and competing methods use matched inputs within each subset.

Table 3. Captured-view selection on Mip-NeRF 360, Train 9, Repair 48, K=3K=3. Each row averages the same full scene set. The test images and optimization seed are fixed; repair-camera identities may change with the training subset.

Original captured subset

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
Train-only17.2100.47760.383017.2170.47780.3823
Global 1.017.6150.48240.400618.0160.49310.4542
Global 0.517.7320.49100.402918.0810.49960.4454
Global 0.317.7790.49570.405018.1230.50450.4370
FixTrust-GS18.0580.50660.402018.1950.51270.4179

Alternative subset: selection seed 137

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
Train-only17.3950.48290.377817.3950.48290.3778
Global 1.017.8410.48790.398118.2270.49680.4510
Global 0.517.9210.49620.400818.2940.50360.4425
Global 0.317.9930.50090.402518.3170.50780.4350
FixTrust-GS18.2620.51130.399518.4360.51720.4158

Alternative subset: selection seed 271

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
Train-only16.7580.46780.395516.7580.46780.3955
Global 1.017.2310.47320.412217.6480.48500.4623
Global 0.517.3260.48130.414917.7030.49100.4559
Global 0.317.3730.48600.416517.7410.49580.4477
FixTrust-GS17.6480.49720.413517.7890.50360.4289

Table 3 shows higher average PSNR and SSIM than the tested scalar weights for both models under all three selections. Relative to Global 0.3, the PSNR margins range from 0.269–0.279 dB for Difix3D and 0.047–0.119 dB for Diffusion Harmonizer. LPIPS improves over Global 0.3 in each selection, but Difix3D remains worse than Global 1.0 in LPIPS. Thus the margins vary with the captured views, and the gains are not uniform across metrics. We will report both robustness protocols and results with these limitations.

Q3. Matched reliability-aware comparison

We evaluated HAD's released reliability scorer with its preprocessing, threshold 0.9, and three nearest captured-view references. The repaired images, camera splits, saved checkpoint, loss, and continuation schedule are identical; only the supervision weights differ. This compares reliability components within a common pipeline, not the complete HAD reconstruction system. Table 4 shows higher PSNR and SSIM and lower LPIPS for our method with both repair models. We will include this matched comparison and clearly describe its scope.

Table 4. Matched HAD comparison on Mip-NeRF 360, Train 9, Repair 48. HAD uses its reliability scorer in the common training pipeline.

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
HAD reliability mask17.8190.49350.403818.0080.49660.4425
FixTrust-GS18.0580.50660.402018.1950.51270.4179
Reviewer_HvLGOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

FixTrust-GS addresses the reliability of supervision generated by pretrained repair models for sparse-view 3D Gaussian reconstruction. It trains auxiliary Gaussian scenes on complementary subsets of the captured images and renders them at the repair cameras. Each repaired pixel receives a weight based on how much it reduces disagreement with these auxiliary predictions relative to the original rendering. These weights modulate the photometric loss during continued optimization of the main Gaussian scene. Experiments cover three datasets, two repair models, and three input view budgets.

The problem is well motivated, and the controlled experiments and direct analysis of repair reliability provide useful evidence for the approach. However, comparisons with existing confidence methods and recent repair pipelines are missing. Gains over carefully tuned scalar weighting are sometimes small or inconsistent, while the cost of training multiple auxiliary reconstructions is not quantified. The lack of supplementary videos also limits assessment of visual consistency across novel views. My preliminary recommendation is weak rejection.

Critical strengths
  1. A meaningful problem and coherent design. Evaluating whether a repair improves agreement with captured-scene evidence directly addresses the risk of incorporating hallucinated details into a persistent 3D representation. The method can use different pretrained repair models through their RGB outputs.

  2. Careful controls and extensive evaluation. The compared integration strategies share the initial reconstruction, repair outputs, cameras, and continuation protocol. The appendix explores global weights from 0.1 to 1.0 and several repair-view counts, providing a substantial assessment of simpler alternatives.

  3. Direct evidence for the reliability weights. The ground-truth diagnostics show that higher weights correlate with more beneficial repairs. The analysis also compares against uniform allocation with the same total weight, supporting the value of selective weighting beyond reducing overall repair strength.

Critical weaknesses
  1. Missing comparisons with representative repair methods. Experiments mainly compare against training without repairs and global scalar weighting, without comparisons to existing confidence strategies or recent repair pipelines such as ArtiFixer3D. This leaves the advantages of both the proposed score and the complete reconstruction pipeline insufficiently established.

  2. The gains over tuned scalar weighting are not consistently substantial. In Table 13, the Deep Blending three-view setting gains only 0.04 dB over the best scalar PSNR, with slightly worse SSIM and LPIPS. Table 1 also shows settings where scalar weighting or training without repairs remains competitive or better.

  3. The additional reconstruction cost is not quantified. The default configuration trains three auxiliary models for 3,000 steps each, but total runtime and peak memory are not reported. Smaller final Gaussian representations do not establish whether this additional computation is worthwhile.

  4. Limited qualitative evidence across novel views. No supplementary rendering videos are provided, making it difficult to assess whether the repaired scenes remain visually consistent along camera trajectories and whether flickering or floating artifacts persist.

Questions for authors
  1. Could the authors compare against a representative uncertainty-based or reprojection-based reliability strategy using the same cached repairs and continuation protocol? A comparison with a recent repair pipeline such as ArtiFixer3D under matched input-view and evaluation protocols would also help establish the practical value of the approach.

  2. What are the end-to-end runtime and peak-memory costs, including the auxiliary reconstructions, repair inference, and evidence rendering? How does the resulting quality compare with simpler alternatives under a similar computational budget?

  3. What conditions determine whether the proposed weighting offers a meaningful advantage over a tuned scalar weight, particularly in the Deep Blending and LLFF settings where the gains are limited or mixed?

4 comments
AuthorsOpenReview ↗

Comment 1 — Reliability methods and repair pipelines

We appreciate the reviewer's thorough assessment of comparative performance, computational cost, and consistency across novel views. The suggestions guided the additional comparisons and cost measurements reported below.

In result tables, bold marks the best reported value per metric within each comparison, including ties.

W1. Comparisons with confidence and repair methods

We have completed two complementary comparisons on Mip-NeRF 360, Train 9, Repair 48, using the original input and test splits.

Reliability scoring. We apply HAD's released scorer, preprocessing, three nearest captured-view references, and threshold 0.90.9 to the same cached repairs. The checkpoint, continuation schedule, and pixel loss remain fixed; only the reliability weights change. This tests HAD's supervision component, rather than its full reconstruction pipeline. Table 1 shows higher PSNR and SSIM and lower LPIPS for FixTrust-GS with both repair models.

Table 1. Matched reliability-score comparison on Mip-NeRF 360, Train 9, Repair 48. Only the repair-supervision weights change.

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
HAD reliability mask17.8190.49350.403818.0080.49660.4425
FixTrust-GS18.0580.50660.402018.1950.51270.4179

Complete repair pipeline. We also ran the released ArtiFixer3D 14B model with its native 3DGUT renderer and training objective at our image resolution, adapting the schedule to 3k initialization and 9k total steps. Table 2 reports the resulting 3D reconstruction, without ArtiFixer3D+ postprocessing. ArtiFixer3D has lower LPIPS, while FixTrust-GS with Diffusion Harmonizer has higher PSNR and SSIM. We will report these mixed results and the protocol differences. The adapted schedule is shorter than the authors' default setup; these results do not establish a ranking under their default operating conditions.

Table 2. Complete-pipeline results under matched Mip-NeRF 360 input/test views (Train 9, Repair 48). ArtiFixer3D retains its native renderer and objective; this comparison does not isolate the weighting rule.

PipelinePSNR ↑SSIM ↑LPIPS ↓
ArtiFixer3D (14B, adapted schedule)18.1440.49020.3696
FixTrust-GS + Difix3D18.0580.50660.4020
FixTrust-GS + Diffusion Harmonizer18.1950.51270.4179

Q1. Reliability strategies and the scope of comparison

FixTrust-GS controls how repaired RGB images supervise the reconstructed scene; it does not generate the repairs. Reconstruction uncertainty indicates where the current scene is uncertain, but does not by itself establish whether a particular repair improves that scene. Reprojection checks cross-view agreement and additionally requires depth, visibility, and occlusion handling. Our score instead compares the initial and repaired pixels against the same scene-specific predictions.

These differences motivated the original controlled integration study, but do not make the requested comparisons unnecessary. Table 1 now provides a released confidence scorer under the identical cached-repair protocol. We have not completed a separate uncertainty or reprojection baseline under this protocol and will state this limitation. ArtiFixer3D generates its own supervision, so Table 2 evaluates it separately as a complete pipeline, with the renderer and schedule differences disclosed.

AuthorsOpenReview ↗

Comment 2 — Computational cost and matched-budget results

W3. Reconstruction time and peak memory

We agree that fewer final Gaussians do not establish lower total computation. We measured the complete reconstruction on one A100 80 GB per run, with all three auxiliary models trained sequentially and charged in full. Table 3 includes model loading, worker setup, image/weight export, and checkpoint I/O. The measurement starts from posed images and the supplied SfM initialization and ends at the final Gaussian checkpoint; camera calibration, downloads, and one-time compilation are outside this common boundary. Offline test evaluation adds 11.0/10.6 s for Difix3D/Harmonizer and is reported separately. NVML samples total device memory every 50 ms; peaks are maximized across sequential stages, not added.

Table 3. Full-method cost on one A100 80 GB, Mip-NeRF 360, Train 9, Repair 48, K=3K=3. Times are scene means in seconds; peak memory is the maximum across scenes. Indented timings are included in their parent stages.

StageDifix3DDiffusion Harmonizer
Initial reconstruction (3k steps)38.838.8
All auxiliary reconstructions (three ×\times 3k)118.5118.5
Initial rendering and repair stage35.651.2
 Repair inference (included above)11.26.4
Evidence rendering and weight stage60.568.8
 Evidence rendering (included above)12.212.5
Main continuation (3k to 9k)73.472.9
Total reconstruction time326.8350.2
Peak device memory (GiB)10.8310.26

The auxiliary and evidence/weight stages account for approximately 55%/53% of the measured Difix3D/Harmonizer runtime. This overhead is substantial. We will add the full breakdown and the matched-budget quality comparison in our response to Question 2 to distinguish representation size from reconstruction cost.

Q2. Quality under a comparable compute budget

For each scene and repair model, we give the baselines the full measured FixTrust-GS budget, including auxiliary training and evidence construction. Train-only spends the remaining time after initialization on captured-image training; scalar baselines additionally pay for repair generation and use the remaining time for continuation. They share the newly measured initial checkpoint and, where applicable, the same repairs. The learning-rate decay is stretched over the available time while preserving its endpoints; no test-based tuning is used. Every run falls within the predeclared 5% tolerance. Table 4 uses fresh end-to-end runs, so its values differ from the original fixed-step results.

Table 4. Quality at matched complete-pipeline wall-time budgets, using fresh end-to-end runs on Mip-NeRF 360, Train 9, Repair 48. Time and quality are scene means; memory is the largest scene peak. All tested budget baselines are shown.

Difix3D

SupervisionTime (s)Peak (GiB)PSNR ↑SSIM ↑LPIPS ↓
Train-only326.010.0817.1570.46420.3648
Global 1.0326.113.2017.8170.48200.3832
Global 0.5326.013.2117.9190.49100.3835
Global 0.3326.011.9217.9990.49760.3830
FixTrust-GS326.810.8318.1180.50780.3994

Diffusion Harmonizer

SupervisionTime (s)Peak (GiB)PSNR ↑SSIM ↑LPIPS ↓
Train-only349.46.9417.1270.46270.3654
Global 1.0349.410.2618.2820.49900.4443
Global 0.5349.312.5218.3680.50600.4339
Global 0.3349.410.2618.4040.51060.4245
FixTrust-GS350.210.2618.2970.51490.4139

For Difix3D, FixTrust-GS has the highest PSNR and SSIM among these alternatives, but worse LPIPS. For Harmonizer, it has the highest SSIM and improves LPIPS over scalar repair weighting, while Global 0.3/0.5 gives higher PSNR and Train-only has lower LPIPS. We will report these trade-offs explicitly; the extra computation does not improve every metric under a matched budget.

AuthorsOpenReview ↗

Comment 3 — Mixed gains, limitations, and qualitative evidence

W2. Limited or mixed gains over scalar weights

The gains are indeed not uniformly substantial. In the cited Table 13 setting (Difix3D, Deep Blending, Train 3, Repair 48), FixTrust-GS achieves 15.27/0.661/0.651 in PSNR/SSIM/LPIPS, compared with 15.23/0.662/0.645 for Global 0.1. This 0.04 dB gain comes with worse SSIM and LPIPS. A low scalar weight can already suppress much of the repair error; spatial selection adds less when reliable improvements are scarce or auxiliary evidence is weak. In addition, our photometric objective does not directly optimize LPIPS. We will emphasize these trade-offs and the competitive Train-only results on LLFF in the main-results discussion and conclusion.

Q3. When does spatial weighting help?

We expect an advantage when useful and harmful repairs coexist within views, the captured images constrain their distinction, and the accepted corrections provide information missing from the current scene. A scalar must attenuate both kinds of content together, whereas a reliable spatial map can separate them.

The benefit can shrink when a small scalar already suppresses most harmful edits, repairs add little new information, or auxiliary models share errors. With only three Deep Blending inputs, each auxiliary model sees just two captured views, which can make rejection overly conservative or acceptance incorrect. LLFF's forward-facing capture can leave less room for useful corrections when Train-only already reconstructs the observed content well. Under fixed-HW normalization, low mean weights also weaken the overall repair update. These are plausible mechanisms, not established dataset-wide causes. We will add this qualified explanation and retain the mixed Deep Blending and LLFF results as limitations of the method.

W4. Novel-view rendering videos

Individual images cannot adequately reveal the view-dependent artifacts raised by the reviewer. We will add rendering videos to our GitHub repository and provide anonymous access for review. The videos will compare Train-only, uniform weighting, and FixTrust-GS along identical camera trajectories, at matched resolution and playback speed, with close-ups of floating artifacts and view-to-view changes. We will include remaining artifacts as well as improvements and link the videos from the supplement.

I maintain my original assessment. The authors have conducted careful and extensive experiments, which is why I initially rated the paper as Weak Reject rather than Reject. However, I feel that much of the work and experimental evaluation focuses on tuning model parameters and configurations, rather than investigating a clearly defined research question.

Additionally, although the authors mentioned that they would provide qualitative videos in the anonymous repository, I was unable to find them. Without these videos or other convincing qualitative evaluations, it remains difficult to assess the visual quality of the reconstructed dynamics.

Overall, while I appreciate the authors' efforts and the extensive experimental evaluation, my main concerns remain unresolved. I therefore maintain my rating of Weak Reject.

Reviewer_qieeOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FixTrust-GS addresses an important existing problem of unreliable diffusion-based supervision for sparse-view 3D Gaussian Splatting (3DGS). Sparse-view reconstruction is severely underconstrained, and diffusion models can repair rendering artifacts but also introduce hallucinated details that become incorporated into the 3DGS training loop. The paper proposes pixel-wise reliability estimation to selectively use these repaired novel views during continued training of 3DGS model. It trains complementary auxiliary 3DGS models on pose-balanced subsets of the real images. At each repair viewpoint, it compares the original and repaired pixels against the same auxiliary predictions. The positive relative reduction in disagreement determines each pixel’s weight in the repair-supervision loss, aiming to retain useful corrections while suppressing unsupported corrections.

The method is evaluated using Difix3D and DiffusionHarmonizer on benchmarks datasets such as Mip-NeRF 360, LLFF, and Deep Blending with 3, 6, and 9 training views. Results generally show improved PSNR and SSIM over uniform repair weighting, often with fewer Gaussians, although improvements are not universal. The paper also includes few qualitative comparisons, and ablations on fold count and loss normalization.

Critical strengths

The paper tries to address an important existing research problem of interest to the community: how to reliably incorporate diffusion repaired images into the 3DGS training loop in sparse settings.

Critical weaknesses
  1. The paper is difficult to follow in places because of unclear terminology and insufficient guidance on which regions or differences readers should examine in the qualitative figures. Clearer explanations, annotated examples, and more informative captions would improve readability. Overall, the writing and visual presentation need substantial improvement to meet the clarity expected of an ICLR submission. Specific examples include:

Line 092: “At each repair camera, their predictions measure relative error reduction from the sealed to the repaired pixel.” The meaning of “sealed” is unclear at this point.

Line 147: “…the resulting reliability map weights continued training from the sealed main state.” The same ambiguity recurs: does “sealed main state” refer to a saved model and optimizer checkpoint, or to a model kept frozen during training? This distinction should be made explicit.

Lines 154–155: The intended meaning becomes clearer when the authors introduce “the sealed reconstruction θ0\theta_0 and its complete optimization state.” However, the term has already appeared several times. It should be defined at first use and replaced with, or clearly linked to, familiar checkpoint terminology.

Figure 4: The figure is not sufficiently self-contained. In particular, panel (d), “Binned utility,” lacks a legend or colorbar explaining the cell colors, and the two numbers within each cell are not clearly identified in the panel or caption. Although the caption directs readers to Appendix C.2.1, that section provides some details but without explaining the color encoding. These details should be stated directly in the figure or its caption.

  1. The results are shown on difix3d and DiffusionHarmonizer, but comparisons with existing methods that explicitly address hallucinations and unreliable repair supervision are missing. Several closely related methods are discussed in the related work but not compared experimentally. For example, HAD [1.] predicts hallucination scores and uses them to control repaired-image supervision during 3DGS training. A direct comparison would help establish the advantage of the proposed method. Furthermore, the evaluation uses only the non-temporal, render-only version of DiffusionHarmonizer, without clearly explaining this choice. Temporal conditioning could improve cross-frame consistency and reduce inconsistent repairs, although consistency alone is not sufficient for scene fidelity. Evaluating the temporal version, or explaining why it is unsuitable for the experimental setting, would help readers to understand whether the proposed method is still better when the repair model already incorporates such consistency.

  2. Why should three auxiliary models trained with fewer images provide sufficiently reliable evidence to evaluate/score a repair of the main model trained with all images, especially in sparse setting? The paper demonstrates empirical correlation with repair quality, but does not discuss why these weaker models should be correct. In short, the paper has not sufficiently established when predictions from weaker, correlated models are reliable indicators of repair improvement, especially in sparse settings.

  3. Several performance differences between methods in Table 1 and Table 2 are small. Repeating the experiments with multiple random seeds (at least three runs) and reporting means and standard deviations, would help establish whether the improvements are robust and statistically significant.

[1.] HAD: Hallucination-Aware Diffusion Priors for 3D Reconstruction (https://arxiv.org/abs/2605.16873)

Questions for authors

Line 021: "This relative agreement improvement yields dense weights for continued 3DGS optimization" ==> The phrase “dense weights” could be clearer. Rephrasing this as “per-pixel reliability weights that control each repaired pixel’s contribution to the training loss” would make the intended meaning more clear.

3 comments
AuthorsOpenReview ↗

Comment 1 — Writing and terminology

We thank the reviewer for the careful reading and specific recommendations on terminology, auxiliary-model reliability, and experimental validation. These comments helped us clarify the presentation and strengthen the supporting evidence.

In result tables, bold marks the best reported value per metric within each comparison, including ties.

W1. Clarity and visual presentation

We have improved the manuscript's writing and presentation in response to these comments, with the following revisions.

  1. Checkpoint terminology. At the original Line 092, “sealed” is replaced by “initial render.” Sections 3.1 and 3.4 now use “saved pre-repair checkpoint”: the main 3DGS parameters, optimizer state, and densification state. The main model resumes training from this checkpoint; the repair model, auxiliary models, repaired images, and weights remain fixed. Algorithm 1 and the figure labels have been updated accordingly.

  2. Figure 4(d). We added a colorbar. Colors and upper entries show the percentage of pixels improved by repair; lower entries show mean RGB L1 error reduction (10−210^{-2} units; positive is better). Rows represent Train 3/6/9 and columns represent weight bins, with equal averaging across scenes. The caption and Appendix C.2.1 specify both quantities and the color scale.

  3. Qualitative comparisons. Figure 4 and Appendix C.4 now use fewer, larger examples, placing magnified details beside each full image. Colored boxes identify identical regions across methods; PSNR refers to full images. The captions identify the details to inspect.

Q1. Meaning of “dense weights”

Following the suggested wording, we have replaced this phrase in the abstract with “per-pixel reliability weights that control each repaired pixel's contribution to the training loss.” The contribution summary and conclusion use consistent terminology.

AuthorsOpenReview ↗

Comment 2 — HAD comparison and temporal repair

W2. HAD and temporal repair

Comparison with uniform weighting alone is insufficient to address this concern. We completed a matched HAD comparison on Mip-NeRF 360, Train 9, Repair 48, using its released scorer and preprocessing, threshold 0.90.9, and three nearest captured-view references. Repairs, checkpoints, camera splits, training schedule, and loss are fixed; only the supervision weights change. This evaluates HAD's reliability component, not its complete reconstruction pipeline.

Table 1 shows PSNR gains of 0.239/0.187 dB over HAD for Difix3D/Diffusion Harmonizer, respectively, with higher SSIM and lower LPIPS. We will add this comparison and its scope to the experiments.

Table 1. Matched HAD comparison on Mip-NeRF 360, Train 9, Repair 48. HAD uses its reliability scorer in the common training pipeline.

SupervisionDifix3D — PSNR ↑Difix3D — SSIM ↑Difix3D — LPIPS ↓Diffusion Harmonizer — PSNR ↑Diffusion Harmonizer — SSIM ↑Diffusion Harmonizer — LPIPS ↓
HAD reliability mask17.8190.49350.403818.0080.49660.4425
FixTrust-GS18.0580.50660.402018.1950.51270.4179

Temporal Diffusion Harmonizer. The original experiments used independent single-frame repair, which did not address temporal conditioning. We have now evaluated the released autoregressive conditioning on four preceding repaired frames, with non-temporal warmup. Cameras follow original image order; the checkpoint, inference settings, camera splits, auxiliary predictions, and training schedule remain matched. We recompute weights from the temporal outputs. Neither weighting nor temporal context uses repair-view or test-view ground truth.

With temporal conditioning, FixTrust-GS improves all three metrics over every tested uniform weight (Table 2). Its temporal variant performs slightly worse than its non-temporal variant on these averages. We will add this configuration and comparison.

Table 2. Matched temporal-conditioning comparison on Mip-NeRF 360, Train 9, Repair 48. All rows cover the same complete scene set.

SupervisionNon-temporal — PSNR ↑Non-temporal — SSIM ↑Non-temporal — LPIPS ↓Temporal — PSNR ↑Temporal — SSIM ↑Temporal — LPIPS ↓
Global 1.018.0160.49310.454217.8380.49150.4508
Global 0.918.0400.49430.451617.8600.49270.4498
Global 0.818.0600.49580.450117.8720.49380.4493
Global 0.718.0600.49680.449417.8900.49530.4472
Global 0.618.0820.49830.447517.9070.49680.4454
Global 0.518.0810.49960.445417.9150.49850.4430
Global 0.418.1060.50200.441217.9670.50090.4401
Global 0.318.1230.50450.437017.9840.50310.4364
Global 0.218.1330.50770.430418.0120.50600.4309
Global 0.118.1070.51000.421918.0070.50820.4235
FixTrust-GS18.1950.51270.417918.1430.51170.4189
AuthorsOpenReview ↗

Comment 3 — Auxiliary evidence and seed robustness

W3. Weaker, correlated auxiliary models

We agree that these predictions are not automatically reliable. They provide captured-scene evidence for a paired comparison, rather than ground-truth estimates. Both candidates are compared with the same auxiliary renders, trained only on captured photographs. Their usefulness depends on locally distinguishing the candidates, even when overall reconstruction accuracy is lower; cross-fitting does not make their errors independent.

Let S,F,IgtS,F,I^{\mathrm{gt}} denote initial, repaired, and ground-truth pixels, RkR_k an auxiliary prediction, and ρ\rho the channel-averaged L1 distance. The reverse triangle inequality gives

g_{\mathrm{gt}}&=\rho(S,I^{\mathrm{gt}})-\rho(F,I^{\mathrm{gt}}), &\widehat g&=\frac{1}{K}\sum_k[\rho(S,R_k)-\rho(F,R_k)],\\\\ \eta&=\frac{1}{K}\sum_k\rho(R_k,I^{\mathrm{gt}}), &|\widehat g-g_{\mathrm{gt}}|&\leq2\eta. \end{aligned}$$ Thus $\widehat g>2\eta$ implies true improvement, but the unknown ground truth prevents using this bound as a training-time certificate. Shared errors can accept incorrect repairs or reject useful ones. Increasing $K$ alone does not eliminate these shared errors. The existing ground-truth diagnostics in Figure 4 empirically support the criterion, but do not guarantee correctness for every pixel. We will clarify these reliability conditions and the limitation of shared errors in the method discussion. ### W4. Repeated runs and variability To assess robustness beyond a single run, we completed seeds 42/43/44 for both repair models at Train 9, Repair 48, $K=3$, on Bicycle, Bonsai, Counter, Garden, Kitchen, Room, and Stump. Camera splits remain fixed; each seed retrains the main and auxiliary models and regenerates repaired supervision. We average scenes equally within each seed, then compute the mean and sample standard deviation across seeds. Table 3 includes every tested uniform weight. FixTrust-GS has higher mean PSNR and SSIM and lower mean LPIPS than every tested uniform weight for both models. We will add these statistics and their aggregation protocol, clearly separating this subset from full-dataset results. They do not establish statistical significance for every setting in the original tables. **Table 3.** Mean $\pm$ sample standard deviation over seeds 42/43/44, Train 9, Repair 48, $K=3$, on the stated Mip-NeRF 360 subset. | Supervision | Difix3D — PSNR ↑ | Difix3D — SSIM ↑ | Difix3D — LPIPS ↓ | Diffusion Harmonizer — PSNR ↑ | Diffusion Harmonizer — SSIM ↑ | Diffusion Harmonizer — LPIPS ↓ | | --- | ---: | ---: | ---: | ---: | ---: | ---: | | Train-only | 18.335 ± 0.081 | 0.5291 ± 0.0015 | **0.3306 ± 0.0021** | 18.325 ± 0.111 | 0.5286 ± 0.0020 | **0.3310 ± 0.0021** | | Global 1.0 | 18.687 ± 0.054 | 0.5313 ± 0.0002 | 0.3478 ± 0.0014 | 19.108 ± 0.041 | 0.5414 ± 0.0005 | 0.3982 ± 0.0027 | | Global 0.9 | 18.710 ± 0.031 | 0.5329 ± 0.0006 | 0.3477 ± 0.0010 | 19.111 ± 0.028 | 0.5424 ± 0.0003 | 0.3966 ± 0.0018 | | Global 0.8 | 18.742 ± 0.040 | 0.5346 ± 0.0003 | 0.3477 ± 0.0010 | 19.128 ± 0.024 | 0.5435 ± 0.0003 | 0.3954 ± 0.0014 | | Global 0.7 | 18.755 ± 0.047 | 0.5364 ± 0.0001 | 0.3480 ± 0.0011 | 19.140 ± 0.026 | 0.5447 ± 0.0006 | 0.3941 ± 0.0019 | | Global 0.6 | 18.781 ± 0.036 | 0.5381 ± 0.0004 | 0.3483 ± 0.0009 | 19.163 ± 0.025 | 0.5466 ± 0.0004 | 0.3916 ± 0.0019 | | Global 0.5 | 18.807 ± 0.043 | 0.5403 ± 0.0003 | 0.3488 ± 0.0008 | 19.182 ± 0.034 | 0.5484 ± 0.0004 | 0.3890 ± 0.0024 | | Global 0.4 | 18.832 ± 0.034 | 0.5428 ± 0.0005 | 0.3487 ± 0.0010 | 19.204 ± 0.032 | 0.5508 ± 0.0005 | 0.3850 ± 0.0018 | | Global 0.3 | 18.855 ± 0.058 | 0.5453 ± 0.0006 | 0.3494 ± 0.0007 | 19.224 ± 0.037 | 0.5537 ± 0.0002 | 0.3802 ± 0.0013 | | Global 0.2 | 18.902 ± 0.046 | 0.5488 ± 0.0003 | 0.3487 ± 0.0012 | 19.229 ± 0.012 | 0.5569 ± 0.0008 | 0.3737 ± 0.0019 | | Global 0.1 | 18.906 ± 0.041 | 0.5510 ± 0.0007 | 0.3496 ± 0.0006 | 19.214 ± 0.019 | 0.5598 ± 0.0006 | 0.3645 ± 0.0018 | | FixTrust-GS | **19.163 ± 0.066** | **0.5573 ± 0.0010** | 0.3430 ± 0.0015 | **19.317 ± 0.032** | **0.5631 ± 0.0003** | 0.3596 ± 0.0012 |

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.