acceptodds
Under review as a conference paper at ICLR 2027

Bi-FlowGS: Bridging Generative View Completion and Gaussian Geometry Through Bidirectional Flow Co-Refinement

Abstract

Sparse-view 3D scene reconstruction with 3D Gaussian Splatting (3DGS) is inherently underconstrained. Plausible renderings can also coexist with erroneous Gaussian geometry, as errors in positions or depths may be concealed by opacity, scale, and appearance; we term this failure mode Geometry Cheating. Existing regularization methods constrain geometry but remain limited to observed views, while video-diffusion-based methods complete unseen views yet mainly use them as RGB pseudo-supervision, underusing motion and temporal priors and lacking explicit geometry supervision. We present Bi-FlowGS, which capitalizes on optical flow to bridge generative view completion and Gaussian geometry regularization. Our plug-and-play Video-to-Geometry Flow Distillation (V2G) distills temporal correspondence priors from restored videos into Gaussian geometry to alleviate Geometry Cheating. Conversely, Geometry-to-Video Flow-Guided Restoration (G2V) uses the current 3DGS geometry to guide temporally consistent video restoration, providing more reliable generative supervision. Together, V2G and G2V form an implicit bidirectional co-refinement process, enabling restored videos and the optimized 3DGS scene to iteratively improve each other. Experiments demonstrate improved rendering quality and geometric consistency across wide-baseline and unbounded 360° benchmarks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Sparse-View Reconstruction, Optical Flow, Novel View Synthesis, Geometry Optimization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WQuLOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bi-FlowGS to address insufficient geometric constraints in sparse-view 3DGS. It introduces V2G to refine Gaussian geometry using optical flow from restored videos and G2V to improve video restoration through geometry-guided flow, jointly enhancing reconstruction and rendering quality.

Critical strengths
  1. Effective bidirectional refinement framework. The proposed V2G and G2V modules establish a bidirectional connection between generative view restoration and Gaussian geometry optimization, improving both rendering quality and geometric consistency.

  2. Extensive experimental validation. The method demonstrates consistent improvements over several baselines across multiple datasets, supported by component ablations and geometric evaluations.

Critical weaknesses

W1. Limited comparison with recent SOTA methods. The paper omits comparisons with recent methods such as ArtiFixer[1], which also leverages video diffusion for sparse-view 3DGS reconstruction. This limits the assessment of its competitiveness against current state-of-the-art approaches.

W2. Unclear presentation of core contributions. The pipeline figure contains excessive details of existing architectures, such as attention mechanisms, while insufficiently highlighting the proposed V2G and G2V modules, making the actual contributions difficult to identify.

W3. Limited robustness analysis of generated optical flow. V2G relies on optical flow from generated videos, which may contain geometrically inconsistent or hallucinated content. Despite reliability filtering, the method's robustness to such errors remains insufficiently validated.

[1] R. de Lutio, T. Fischer, Y.-Y. Chang, Y. Zhang, J. Z. Wu, X. Ren, T. Shen, K. Tothova, Z. Gojcic, and H. Turki, “ArtiFixer: Enhancing and Extending 3D Reconstruction with Auto-Regressive Diffusion Models,” in Proc. ACM SIGGRAPH, 2026.

Questions for authors

Q1. Comparison with recent SOTA methods. Could the authors provide comparisons with recent methods such as ArtiFixer under consistent evaluation settings to better demonstrate the competitiveness of Bi-FlowGS?

Q2. Robustness to generated flow errors. How does V2G handle geometrically inconsistent optical flow caused by hallucinated content in generated videos? Can the authors provide further robustness analysis?

8 comments
AuthorsOpenReview ↗

Response to Reviewer WQuL

We thank the reviewer for recognizing the bidirectional co-optimization framework and effectiveness of our method, as well as the comprehensiveness of our experiments. Below, we provide additional evidence and analyses addressing the weaknesses you raised.

Weakness1&Question1: Insufficient comparisons with the latest SOTA methods.

Answer:

We thank the reviewer for this helpful suggestion. To address your concern, we have added comparisons with ArtiFixer and plug-and-play experiments, as detailed below:

(1) Reconstruction comparison with ArtiFixer. We compare Bi-FlowGS with ArtiFixer on CO3D under the 3-, 6-, and 9-view input settings, strictly ensuring identical input views and test splits for both methods to enable a fair comparison. The results are presented in the following table.

Comparison with ArtiFixer on CO3D:

MethodPSNR ↑ (3-view)PSNR ↑ (6-view)PSNR ↑ (9-view)SSIM ↑ (3-view)SSIM ↑ (6-view)SSIM ↑ (9-view)LPIPS ↓ (3-view)LPIPS ↓ (6-view)LPIPS ↓ (9-view)
ArtiFixer16.0319.2920.720.6020.6460.6800.4610.4010.365
Bi-FlowGS (Ours)15.9219.3320.780.5950.6500.6820.4740.3980.359

As shown in the results, ArtiFixer achieves stronger reconstruction performance under the extremely sparse 3-view input setting, possibly because the initial reconstruction is less complete, limiting the effectiveness of flow guidance and geometric supervision in Bi-FlowGS. As the number of input views increases, the advantages of Bi-FlowGS become more pronounced: our method improves all three metrics under the 6- and 9-view settings. These results demonstrate the competitiveness of Bi-FlowGS.

(2) Plug-and-play effectiveness of V2G on ArtiFixer. Following the plug-and-play evaluation protocol in Table 3 of the original manuscript, we further integrate V2G into ArtiFixer and conduct experiments on CO3D under the 6-view setting. The results are shown below. Geo. Avg follows the definition in Table 3 and denotes the average relative improvement across eight geometry metrics.

Plug-and-play results of V2G on ArtiFixer:

MethodPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓SqRel ↓RMSE ↓RMSE-log ↓MedianRel ↓δ₁ ↑δ₂ ↑δ₃ ↑Geo. Avg ↑
ArtiFixer19.290.6460.4010.10790.02300.14990.16700.07900.83280.98150.9969—
ArtiFixer + V2G19.420.6490.3970.04670.01040.07380.08090.03260.86120.98420.9972+34.53%

With V2G, all rendering metrics improve, with an average gain of 34.53% across the geometry metrics. These results extend the plug-and-play evaluation to ArtiFixer, supporting the applicability of V2G to another recent reconstruction framework.

AuthorsOpenReview ↗

Response to Reviewer WQuL

Weakness2: Unclear presentation of the core contributions and redundant presentation of attention-mechanism details.

Answer:

We thank the reviewer for this constructive suggestion. To help you understand our pipeline diagram more clearly, we have improved its structure and explicitly labeled V2G and G2V to highlight the core contributions. The updated diagram appears in Figure 2 of the revised manuscript.

Regarding the attention-mechanism details, Flow-Biased Local Cross-Attention is a new module that we have added to the CogVideoX backbone. It uses flow-derived attention biases to transfer spatially aligned, fine-grained features from the first and last real reference images to each generated frame. Complementing geometry-aware motion guidance, it enriches texture details while maintaining motion consistency. We therefore retain this module in the pipeline diagram to illustrate how G2V jointly preserves motion consistency and local appearance fidelity.


Question2.1: How does V2G handle errors when hallucinated content in generated videos produces geometrically inconsistent flow?

Answer:

a. V2G incorporates a teacher-flow filtering mechanism that considers both temporal consistency and geometric validity. Specifically, the teacher-flow reliability term excludes matches with invalid warps, excessive motion, or large forward–backward cycle errors; the geometric reprojection mask further excludes occluded or geometrically inconsistent correspondences. The detailed design is provided in Appendix E.4.

b. Ablation Study Demonstrates the Effectiveness of the Flow Filtering Mechanism.

We conducted an ablation study of the filtering mechanism in the paper, providing preliminary evidence of its effectiveness. The following results are taken from Appendix E.4.

Ablation study of the flow filtering mechanism:

Teacher Flow ReliabilityGeometric Reprojection MaskPSNR ↑SSIM ↑LPIPS ↓
✓✓15.6430.5020.468
×✓15.4420.4990.469
✓×15.4540.4980.470
××15.4800.4990.468

Removing the V2G flow filtering mechanism clearly degrades the results, demonstrating the effectiveness of teacher-flow filtering.


Question2.2&Weakness3: Insufficient robustness analysis regarding flow errors.

Answer:

We sincerely thank the reviewer for this valuable suggestion. In response, we have conducted additional ablation experiments to evaluate the robustness of our method to erroneous optical flow. The experimental results and corresponding analysis are presented below.

(1) V2G effectively filters flow across different quality levels. To analyze the robustness of our filtering mechanism to erroneous flow, we categorize flow by quality and examine filtering performance at each quality level. Under the 6-view setting on Tanks and Temples, we randomly select three scenes, generate restored videos along the GT video trajectories using the video diffusion model, and estimate flow from the restored videos using WAFT. Based on LPIPS between the restored and GT videos, we divide restoration quality into Good (LPIPS ≤ 0.3619), Medium (0.3619 < LPIPS ≤ 0.4349), and Poor (LPIPS > 0.4349) groups and compare flow metrics before and after filtering.

The evaluation includes 1,793 valid frame pairs in total.

Flow filtering across different quality levels (units: pixels):

Restoration QualityFrame PairsLPIPS ↓FB Before ↓FB After ↓EPE Before ↓EPE After ↓Fl-all Before ↓Fl-all After ↓Masked Ratio
Good4280.32762.1841.2862.8362.61525.48%13.72%26.20%
Medium9760.40893.0181.6203.4872.82534.72%19.46%41.10%
Poor3890.49323.9962.4304.9643.27750.65%32.18%59.30%
Overall1,7930.40783.0311.7163.6522.87335.97%20.85%41.49%

Filtering reduces FB Error, Geometric EPE, and Fl-all in all three quality groups, while the proportion of filtered matches increases as flow quality deteriorates. These results indicate that V2G is robust to flow of different quality levels. The corresponding experiment has been added to Appendix E.6.

AuthorsOpenReview ↗

Response to Reviewer WQuL

(2) Validating the Effectiveness of Flow Filtering through Reconstruction Comparison with GT Video Flow.

To further validate the effectiveness and robustness of V2G's flow filtering mechanism, we conduct reconstruction experiments on Tanks and Temples under the 6-view setting. We change only the source of teacher flow while keeping all other training procedures identical. Specifically, we feed the GT video frames and the diffusion-restored frames generated along the same GT camera trajectory into the same WAFT model, obtaining GT Video Flow and Ours Flow, respectively, and compare their reconstruction results.

Reconstruction comparison using different teacher-flow sources (one complete video-generation round):

Teacher FlowPSNR ↑SSIM ↑LPIPS ↓Surface F-score (%) ↑AbsRel ↓
GT Video Flow15.920.5570.44614.030.05864
Ours Flow15.830.5560.44911.230.08100

The two teacher-flow sources achieve comparable rendering quality. These results indicate that Ours Flow provides effective supervision and that V2G's flow filtering mechanism helps mitigate, to some extent, the impact of unreliable flow on final reconstruction, further demonstrating its effectiveness and robustness. We have added the corresponding experiment to Appendix E.5.

(3) Failure-case analysis. To analyze extreme teacher-flow failures, we reuse the experiment in (1) and evaluate the same 1,793 valid frame pairs. A frame pair with a mean EPE exceeding 5 pixels is defined as an extreme failure in this analysis. A total of 105 frame pairs (5.86%) meet this criterion, with a mean EPE of 8.141 px. After applying the reliability mask, 80.64% of valid geometric-evaluation pixels are discarded, leaving 45 frame pairs whose mean EPE decreases to 4.353 px. Among these 45 frame pairs, 15 pairs still have a mean EPE above 5 pixels.

Filtering of flow in extreme failure cases:

EvaluationFrame PairsGeometric EPE ↓Masked Ratio (Proportion of Discarded Flow)
Before Mask1058.141 px—
After Mask454.353 px80.64%

Direct inspection of erroneous flow matches that pass filtering reveals why they can pass: paired frames may contain visually consistent content arising from shared hallucinations generated by the model. Consequently, the estimated pixel correspondences satisfy forward–backward consistency while remaining geometrically incorrect. We have also added visualizations of the corresponding flow failure cases in Figure 14 of the appendix.

These results show that the reliability mask removes most unreliable matches in extreme failure cases and reduces the geometric error of retained flow. We have added the corresponding experiment to Appendix E.6.

(4) Artificial perturbation analysis. To further verify robustness, we artificially add noise perturbations to the teacher flow before applying our reliability filtering. We then analyze the resulting reconstruction performance.

For each teacher-flow map generated during iteration, we randomly select a rectangular region covering approximately 10% of the image area. Within this region, we add a uniform offset to all flow vectors. The offset has a randomly sampled direction and a fixed magnitude of either 8 or 16 pixels.

All other reconstruction settings remain unchanged. We conduct reconstruction experiments under the 6-view setting on Tanks and Temples, evaluating four variants corresponding to the two perturbation magnitudes, with the reliability mask either retained or removed.

Reconstruction comparison with and without mask filtering under artificial teacher-flow perturbations:

Experimental VariantPSNR ↑SSIM ↑LPIPS ↓
Ours15.930.5550.437
8 px + Mask15.900.5530.439
8 px, no mask15.760.5470.446
16 px + Mask15.870.5490.441
16 px, no mask15.680.5390.455

With the reliability mask retained, perturbations have only a minor effect on reconstruction, and the resulting reconstruction performance is stronger than without mask filtering. These results demonstrate that our filtering mechanism maintains robust reconstruction performance across different flow quality levels. The corresponding experiment has been added to Appendix E.6.


We sincerely thank you again for your positive assessment and constructive suggestions. We hope that our responses have adequately addressed your concerns. Should you have any further questions or concerns, we would be more than happy to discuss them with you.

Reviewer_BeV2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents Bi-FlowGS, a framework for sparse-view 3D Gaussian Splatting reconstruction that connects generative view completion with geometry optimization through optical flow. Video-to-Geometry Flow Distillation (V2G) aligns geometry-induced flow with teacher flow estimated from restored videos. Geometry-to-Video Flow-Guided Restoration (G2V) uses flow and geometric cues from the current reconstruction, together with reference appearance features, to guide video restoration. The two components form an iterative feedback loop. Experiments demonstrate improvements in novel-view rendering and geometric consistency across multiple benchmarks.

Critical strengths
  1. The experimental evaluation is fairly comprehensive, covering multiple datasets, component ablations, and both rendering and geometry metrics. Overall, the results support the effectiveness of the proposed method.
  2. V2G is evaluated within several existing reconstruction frameworks and generally improves their results. These experiments provide useful evidence of its applicability beyond the complete Bi-FlowGS pipeline.
  3. The paper is clearly written and easy to follow, with a well-organized presentation of the motivation, method, and experiments.
Critical weaknesses
  1. The methodological novelty of V2G appears limited relative to Flow Distillation Sampling (FDS). Both methods use flow estimated by a pretrained matching model to supervise flow derived from rendered depth and camera poses, thereby regularizing Gaussian geometry. V2G differs primarily in obtaining its teacher flow from diffusion-restored videos and incorporating reliability filtering and weighting. In my view, these changes do not provide sufficient methodological novelty to justify presenting V2G as a major contribution of the paper.
  2. The reliability of the teacher flow remains insufficiently examined. G2V-restored videos may contain temporally coherent content that is inconsistent with the geometry of the original scene. Flow estimated from such videos could therefore pass consistency checks while still providing incorrect geometric supervision. Although the downstream results and reliability-mask ablations are encouraging, a more direct analysis of teacher-flow accuracy and failure cases would help establish when this supervision is reliable and whether restoration errors limit the method's effectiveness.
Questions for authors

I am curious whether replacing flow supervision with depth supervision at the restored novel views would yield comparable improvements. For example, teacher depth could be estimated from the G2V-restored frames using Depth Anything V2 and used with an appropriately aligned depth loss. A controlled comparison using the same restored views and optimization budget would help clarify whether the gains from V2G arise specifically from flow-based supervision or, more generally, from introducing geometric supervision at novel views.

8 comments
AuthorsOpenReview ↗

Response to Reviewer BeV2

We sincerely thank the reviewer for recognizing our extensive experiments, affirming our overall approach and V2G's effectiveness, and positively assessing our manuscript's presentation. We address your concerns point by point below.

Weakness1: V2G and FDS share certain methodological similarities, as both use pretrained-matcher flow to supervise geometry-derived flow and regularize Gaussian geometry.

Answer:

We thank the reviewer for this comment. We agree that both V2G and FDS use flow estimated by a pretrained matching model to supervise geometrically derived flow. Beyond this shared mechanism, however, we would like to clarify for you that our method introduces substantive innovations and differs substantially from FDS. We elaborate below:

  1. Connecting generative view completion with explicit geometric constraints through optical flow. FDS-type methods can optimize Gaussian geometry but lack generative completion capabilities, making it difficult for them to independently recover scene content across a wide range of unseen viewpoints. Bi-FlowGS combines the complementary advantages of the two categories of methods to support both scene completion and geometric accuracy.
  2. Designing a video diffusion model that preserves both motion consistency and texture details. Through geometry-aware motion guidance, global appearance cross-attention, and flow-biased local cross-attention, G2V jointly constrains overall motion, global appearance semantics, and local details, providing more effective teacher flow and pseudo-view supervision.
  3. Providing more effective geometric supervision using video priors. V2G distills the temporal motion and implicit geometric priors learned during video diffusion pretraining into Gaussian geometry. In comparison, matching in FDS relies on existing observations and current renderings, and its supervision is limited by missing content and rendering artifacts.
  4. Connecting video restoration and geometry optimization through implicit bidirectional co-refinement. The updated Gaussian scene provides new restoration conditions, and the restored videos in turn update teacher flow and pseudo-view supervision, allowing both components to improve jointly as reconstruction progresses and directly optimizing Gaussian geometry while completing scene content across a wide range of unseen viewpoints.
  5. Greater Practical Value. Bi-FlowGS offers greater practical value than FDS. As demonstrated by the experiments in the paper, it performs strongly across four sparse-view benchmarks and plug-and-play evaluations spanning three categories of recent methods. We further validate its practical advantages over FDS through the experiments presented below.
AuthorsOpenReview ↗

Response to Reviewer BeV2

Below, we further explain to you the contributions of our method and its differences from FDS through experiments.

(1) More robust and effective geometric supervision than FDS.

a.To directly compare their transferability, we add FDS plug-and-play experiments under the 6-view setting on Mip-NeRF 360, following Table 3 of the original manuscript. For each base framework, we retain its original pipeline and optimization settings and separately integrate FDS or V2G for comparison.

Plug-and-play comparison of V2G and FDS:

MethodPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓SqRel ↓RMSE ↓RMSE-log ↓MedianRel ↓δ1 ↑δ2 ↑δ3 ↑Geo. Avg ↑
GenFusion16.800.4270.4940.03010.01000.10640.12400.00700.96030.97660.9843—
GenFusion + FDS16.870.4330.4910.01820.00820.09360.15650.00510.98720.98880.9893+9.37%
GenFusion + V2G16.940.4310.4940.01910.00650.09120.08900.00540.97990.98200.9974+17.39%
GSFixer17.360.4300.4750.08510.03880.25550.17300.03760.88280.96540.9888—
GSFixer + FDS17.140.4370.4740.06920.03690.22740.16440.02160.92970.97320.9861+10.99%
GSFixer + V2G17.430.4320.4680.03180.00780.10690.10160.00900.96860.99330.9946+41.42%

V2G achieves Geo. Avg gains of 17.39% and 41.42% in the two frameworks, exceeding the corresponding FDS gains of 9.37% and 9.35%. FDS improves some geometry metrics but reduces PSNR on GSFixer, further validating the robustness of V2G across different paradigms. These results directly demonstrate the effectiveness of our restored-video-based geometric supervision.

b. Replacing V2G with FDS as geometric supervision within the same framework. To fairly and directly compare the geometric supervision provided by FDS and V2G, we retain G2V and all other reconstruction procedures and optimization settings under the 6-view setting on Tanks and Temples, replacing only V2G with FDS. The following results are reproduced from Figure 8(a) of the original manuscript.

Comparison of V2G and FDS geometric supervision within the same framework:

ConfigurationPSNR ↑SSIM ↑LPIPS ↓Surface F-score (%) ↑
G2V + FDS15.530.5480.4529.21
G2V + V2G (Ours)15.930.5550.43711.62

With the remaining pipeline unchanged, V2G improves both rendering quality and geometric accuracy, directly validating the effectiveness of distilling flow from restored videos for geometric supervision.

(2) Stronger reconstruction capability across a wide range of unseen viewpoints than FDS. To evaluate reconstruction from sparse views across a wide range of viewpoints, we further compare with the original FDS method under the 3-, 6-, and 9-view settings on Mip-NeRF 360:

Sparse-view reconstruction comparison for scene content across a wide range of unseen viewpoints:

MethodPSNR ↑SSIM ↑LPIPS ↓
3-view6-view9-view3-view6-view9-view3-view6-view9-view
FDS13.1314.9616.910.3080.3990.4750.5540.4850.415
Bi-FlowGS (Ours)16.0817.7019.060.3880.4460.5010.5410.4680.413

Bi-FlowGS improves all three rendering metrics under all three input-view settings. These results further demonstrate the practical value of combining generative completion with geometric constraints.

AuthorsOpenReview ↗

Response to Reviewer BeV2

Weakness2.1: Videos restored by G2V may contain temporally coherent content that is inconsistent with the original scene geometry.

Answer:

We thank the reviewer for this important question. We agree that temporal consistency alone does not guarantee geometric accuracy. However, (1) our direct accuracy analysis shows that, despite this issue, flow obtained from restored videos has higher overall quality than flow estimated from rendered images, as well as relatively high 3D geometric accuracy. (2) The failure-case analysis and flow filtering analysis across different quality levels show that, despite the issue raised by the reviewer, our mask can robustly prevent such problematic flow from entering the training process. (3) The flow selection mechanism in V2G considers both temporal and geometric consistency of teacher flow(see Appendix E.6(a)), thereby reducing the impact of unreliable flow on the reconstruction.(4) Reconstruction comparisons using GT videos for supervision show that our video-estimated flow provides effective supervision; therefore, this issue does not introduce substantial reconstruction or geometric deviations.

The detailed results of the experiments are provided below.


Weakness2.2: A more direct analysis of teacher-flow accuracy.

Answer:

Under the 6-view reconstruction setting on Tanks and Temples, we randomly select three scenes and separately evaluate the accuracy of the following three flow sources:

  1. Raw Render Flow: WAFT-estimated flow from Gaussian-rendered frames along the GT camera trajectory.

  2. Restored Video Flow: WAFT-estimated flow from diffusion-restored frames along the same trajectory.

  3. Real-Video Flow: WAFT-estimated flow from corresponding real-video frames along the same trajectory.

The evaluation includes 1,793 valid frame pairs shared by all three flow sources.

Accuracy evaluation of flow from different sources (without WAFT):

Flow SourceFB Error ↓ (Temporal Consistency)EPE ↓(Geometric Consistency)Fl-all ↓ (Outlier Rate)
Raw Render5.193 px4.150 px62.46%
Restored (Ours)3.031 px3.652 px35.97%
GT Video1.230 px2.490 px11.27%

The results show that despite some geometric inconsistencies, restored-video flow substantially outperforms rendered-image flow in both temporal and geometric consistency. Furthermore, the accuracy of our flow is close to that of flow estimated from GT videos, further supporting the inherent reliability of restored-video flow. We also visualize different types of flow, providing further evidence for the reliability of teacher flow. The experiment has been added to Appendix E.5, and the visualizations are presented in Figure 18.


Weakness2.3: Failure-case analysis.

Answer:

To analyze extreme teacher-flow failures, we reuse the experiment in Weakness2.2 and evaluate the same 1,793 valid frame pairs. A frame pair with a mean EPE exceeding 5 pixels is defined as an extreme failure in this analysis. A total of 105 frame pairs (5.86%) meet this criterion, with a mean EPE of 8.141 px. After applying the reliability mask, 80.64% of valid geometric-evaluation pixels are discarded, leaving 45 frame pairs whose mean EPE decreases to 4.353 px. Among these 45 frame pairs, 15 pairs still have a mean EPE above 5 pixels.

Filtering of flow in extreme failure cases:

EvaluationFrame PairsGeometric EPE ↓Masked Ratio (Proportion of Discarded Flow)
Before Mask1058.141 px—
After Mask454.353 px80.64%

Direct inspection of erroneous flow matches that pass filtering reveals why they may evade the filter: the paired frames may contain visually consistent content arising from shared hallucinations. Consequently, the estimated pixel correspondences may satisfy forward–backward consistency while remaining geometrically incorrect. We have also added visualizations of the corresponding erroneous-flow cases to Appendix Figure 14 to validate this hypothesis. These results show that the reliability mask removes a substantial proportion of unreliable matches in extreme failure cases and reduces the geometric error of retained flow. The corresponding experiment has been added to Appendix E.6.

Reviewer_cHL7OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Bi-FlowGS for sparse-view 3dgs reconstruction. The two proposed components, ie, v2g and g2v, are used iteratively to improve reconstruction quality.

Critical strengths

The idea of using optical flow to supervise gaussian geometry is simple and intuitive.

Critical weaknesses
  1. The teaser and pipeline figs are overcrowded, with many redundant illustrations. The overall workflow is difficult to follow.

  2. The method introduction is unclear. It is difficult to connect the equations with the pipeline fig, and the implementation details are not well explained.

  3. V2G and G2V are not truly symmetric. G2V uses optical flow estimated from rendered images rather than directly using the geometry-induced flow defined in V2G.

  4. The reliability of optical flow from generated videos is not well justified. Temporally consistent flow does not necessarily represent correct 3D geometry.

Questions for authors
  1. Why should optical flow estimated from restored videos be more geometrically reliable than flow from the original renderings? Can the authors directly evaluate teacher-flow accuracy?

  2. Can the authors better explain how the equations correspond to the pipeline, especially the flow-guided attention mechanism?

  3. Can the authors provide stronger evidence that bidirectional feedback is necessary, rather than simply combining V2G with video restoration?

7 comments
AuthorsOpenReview ↗

Response to Reviewer cHL7

We sincerely thank the reviewer for the careful review and for identifying the remaining weaknesses of our manuscript. Below, we provide detailed responses to each weakness and question.

Weakness 1: The teaser figure and pipeline diagram are insufficiently clear.

Answer:

We greatly appreciate this helpful and constructive suggestion. Following your recommendation, we have revised the teaser figure and pipeline diagram, reducing redundant illustrations and reorganizing their layouts to improve clarity and readability. The updated figures appear in Figures 1 and 2 of the revised manuscript.


Weakness 2 and Question 2: The method description and pipeline diagram are unclear. Could the authors more clearly explain the correspondence between the equations and the pipeline, particularly the flow-guided attention mechanism?

Answer:

To help you better understand our work, we have improved Figure 2, clearly distinguishing the functions of the individual modules in the pipeline diagram. At the same time, we have improved the structure of Section 3.2 and refined its wording to enhance readability and clarity. We next explain to you how the content of the original manuscript corresponds to Figure 2, as well as the role of flow-guided local cross-attention. The G2V video diffusion model primarily comprises the following three modules:

a. Geometry-aware Motion Guidance (Equation (4) in the revised manuscript): This corresponds to the flow injection and 3D attention modules in the figure. By injecting multiscale flow features before the 3D attention module, it converts cross-view correspondences from the current 3DGS into motion guidance, directly constraining generated videos to maintain temporal motion consistent with the scene.

b. Global Appearance Cross-Attention (Equation (5) in the revised manuscript): This corresponds to the Global cross-attention module following the 3D attention module in the figure. While preserving motion consistency, global cross-attention jointly extracts DINOv2 features from the first and last real reference images, using global appearance and semantic cues to maintain consistent scene appearance.

c.Flow-guided Local Cross-Attention (Equation (6) in the revised manuscript): This corresponds to the separately labeled Flow-Biased Local Attention module in the figure. To supplement global appearance retrieval with fine-grained spatial details, the two local attention operations use flow-derived attention biases to transfer corresponding appearance details from the first or last real reference image to the current generated frame. The original equation is:

= \sum_{r=0}^{1}w_t^r\operatorname{CrossAttention}l \left(\mathbf X{l,t},\mathbf S^r;\mathbf B_{l,t}^{r,F}\right).$$ Here, $\mathbf X_{l,t}$ denotes the normalized video features for frame $t$ from the preceding 3D self-attention module; $\mathbf S^r$ denotes the DINOv2 features of the first/last real reference image; and $\mathbf B_{l,t}^{r,F}$ denotes the flow-derived attention bias encoding spatial correspondences between the current frame and the corresponding endpoint reference image. These three modules provide **complementary constraints on overall motion, global appearance semantics, and local details**, improving appearance consistency and fine-grained fidelity while preserving motion consistency.
AuthorsOpenReview ↗

Response to Reviewer cHL7

Weakness 3: V2G and G2V are not truly symmetric. G2V uses optical flow estimated from rendered images instead of V2G's geometry-induced flow.

Answer:

(1) We would like to clarify for you that we do not claim strict symmetry between the two flow constructions; rather, our formulation is implicit bidirectional co-refinement (Abstract, Section 1, and Section 3.3). However, to help you gain a clearer understanding of our work, we have strengthened this description in the first paragraph of Section 3 (Method) in the revised manuscript.

(2) Furthermore, to better address your concerns, we conducted an ablation study on Mip-NeRF 360 under the 6-view setting, replacing G2V’s FlowCond prior with V2G’s geometry-derived flow and keeping other settings unchanged.

Reconstruction with V2G geometry-derived flow as the G2V prior:

Reduction denotes the relative reduction in AbsRel compared with w/o FlowCond.

G2V PriorPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓Reduction ↑
w/o FlowCond17.500.4390.4670.0295—
Reprojection Flow17.590.4440.4720.020829.49%
Ours17.700.4460.4680.018238.31%

Ours achieves greater gains than reprojection flow over the no-flow baseline, likely because WAFT’s pretrained matching prior reduces reliance on rendered depth and better accommodates inaccurate Gaussian geometry.


Question 3: Could the authors provide stronger evidence for the necessity of bidirectional feedback, beyond simply combining V2G with video restoration?

Answer:

We address this question from the perspectives of theory, existing experimental evidence, and additional experimental evidence.

(1) Necessity of bidirectional feedback. As stated in the introduction of the original manuscript, Bi-FlowGS aims to bridge generative view completion and Gaussian geometry regularization, combining their complementary advantages in novel-view completion and explicit geometric constraints. V2G distills cross-frame correspondences from restored videos into geometric supervision, while G2V guides video restoration using flow and geometric cues from the current Gaussian scene. Through iterative feedback, the two components mutually reinforce each other, jointly supporting scene completion across a wide range of viewpoints and accurate Gaussian geometry.

(2) Practical value. As demonstrated by the experiments in the paper, despite its simple formulation, Bi-FlowGS performs strongly across four sparse-view benchmarks and plug-and-play evaluations spanning three categories of recent methods. We will release the code to support future research.

(3) Existing experiments support bidirectional synergy.

a. Plug-and-play (Table 3): V2G improves geometry across frameworks; with G2V, it achieves the largest combined gains in geometry and rendering, supporting G2V’s targeted design.

b. Ablations and training dynamics (Table 4, Figures 6–7): Combining V2G and FlowCond yields greater gains; visualizations and training curves corroborate their joint benefits.

c. FDS comparison (Figure 8): Replacing V2G with FDS while keeping G2V fixed yields smaller gains, supporting their complementarity.

(4) To better address your concern, we conducted an ablation study of the G2V injection module. We thank the reviewer for this suggestion; the corresponding table and analysis have been added to Table 5 in the revised manuscript. Under the 6-view setting on Mip-NeRF 360, we compare the reconstruction performance of video diffusion models paired with V2G under four conditioning priors. We replace only the G2V conditioning prior and keep all other training settings unchanged. The four priors are: no FlowCond, Plücker ray maps, reprojection flow, and the original WAFT flow prior.

Reconstruction with different G2V conditioning priors:

Reduction denotes the relative reduction in AbsRel compared with w/o FlowCond.

G2V PriorPSNR ↑SSIM ↑LPIPS ↓AbsRel ↓Reduction ↑
w/o FlowCond17.500.4390.4670.0295—
Plücker17.540.4410.4700.024417.29%
Reprojection Flow17.590.4440.4720.020829.49%
Ours17.700.4460.4680.018238.31%

Compared with reprojection flow, WAFT’s pretrained matching prior reduces dependence on rendered depth. Compared with Plücker conditioning, explicit flow provides more direct cross-frame motion correspondences and better meets V2G’s temporal supervision requirements. These results show that the conditioning prior of the video diffusion model affects the final gains under the same V2G supervision, supporting the targeted design of G2V. We also visualize restored-video flow under different conditioning priors, further validating the effectiveness of G2V’s conditioning design; the visualizations have been added as Figure 13 in the appendix.

AuthorsOpenReview ↗

Response to Reviewer cHL7

Weakness 4 & Question 1: Teacher-flow reliability is insufficiently justified: temporal consistency does not guarantee geometric accuracy. Why should restored-video flow be more geometrically reliable than raw-render flow, and can its accuracy be directly evaluated?

Answer:

We thank the reviewer for this important question. Restored-video flow is not necessarily more reliable; however, the video diffusion model completes rendered sequences using pretrained motion and temporal priors, providing cross-frame correspondences beyond the current sparse observations. In contrast, raw-render flow is more directly affected by errors in the current Gaussian geometry and by rendering artifacts. We validate these points experimentally below:

(1) Validating the reliability of restored-video flow through reconstruction.

a. Restored-video flow provides more effective geometric supervision than raw-render flow. Under the 6-view setting on Tanks and Temples, we use the full iterative reconstruction pipeline and change only the source of V2G teacher flow, keeping the remaining pipeline unchanged. Raw Render Flow is estimated from raw rendered images using the frozen WAFT model.

Reconstruction comparison using rendered-image flow as teacher flow:

Teacher FlowPSNR ↑SSIM ↑LPIPS ↓Surface F-score (%)↑AbsRel ↓
Raw Render Flow15.400.5370.46211.490.13607
Ours Flow15.930.5550.43712.910.10611

Surface F-score and AbsRel directly measure geometric reconstruction quality. The consistent improvements in rendering and geometry metrics demonstrate that restored-video flow provides more effective geometric supervision than rendered-image flow.

b. Reconstruction comparison using flow estimated from real videos as teacher supervision. Under the 6-view setting on Tanks and Temples, we further generate restored videos along the GT video trajectories using the video diffusion model and obtain our teacher flow using WAFT. For a fair reconstruction comparison, we change only the teacher flow to flow estimated from the GT videos using the same WAFT model (GT Video Flow).

Reconstruction comparison using GT-image flow as teacher flow (one complete restoration round for final optimization):

Teacher FlowPSNR ↑SSIM ↑LPIPS ↓Surface F-score (%) ↑AbsRel ↓
GT Video Flow15.920.5570.44614.030.05864
Ours Flow15.830.5560.44911.230.08100

The two teacher-flow sources yield similar final rendering quality, supporting the accuracy and a degree of geometric reliability of flow from our restored videos as teacher supervision.

We have added the corresponding experiment to Appendix E.5.

(2) Direct analysis of teacher-flow accuracy and the iterative process.

a. Direct accuracy evaluation of flow from different sources. Building on the reconstruction experiment along GT trajectories described above, we randomly select three scenes and separately evaluate the accuracy of the following three flow sources:

  1. Raw Render Flow: WAFT-estimated flow from Gaussian-rendered frames along the GT camera trajectory.

  2. Restored Video Flow: WAFT-estimated flow from diffusion-restored frames along the same trajectory.

  3. Real-Video Flow: WAFT-estimated flow from corresponding real-video frames along the same trajectory.

The evaluation includes 1,793 valid frame pairs shared by all three flow sources.

Direct flow-accuracy evaluation using WAFT-estimated flow (without reliability filtering):

Flow SourceFB Error ↓ (Temporal Consistency)EPE ↓(Geometric Consistency)Fl-all ↓ (Outlier Rate)
Raw Render5.193 px4.150 px62.46%
Restored (Ours)3.031 px3.652 px35.97%
GT Video1.230 px2.490 px11.27%

The results show that despite some geometric inconsistencies, restored-video flow substantially outperforms rendered-image flow in both temporal and geometric consistency. Comparing the accuracy of our flow with that of flow estimated from GT videos further supports the inherent reliability of restored-video flow.

We have visualized optical flow from different sources, further demonstrating the reliability of teacher flow. The experiment has been added to Appendix E.5, and the visualizations are presented in Figure 18.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary for Submission 3782

Dear Area Chairs and Reviewers,

We sincerely thank all reviewers for their thorough assessment and constructive feedback on our submission. Reviewer WQuL recognized the research significance and methodological effectiveness of our implicit bidirectional co-refinement framework for wide-range unseen-view expansion and Gaussian geometry optimization. Reviewer BeV2 noted V2G’s practical contribution and potential applicability across frameworks and the manuscript’s clarity and organization. Reviewer cHL7 found flow-based geometry supervision intuitive and well motivated. Reviewer WQuL and BeV2 praised the comprehensive experimental coverage and competitive results.

We have addressed all reviewer questions point by point; key clarifications and the additional experiments are summarized below.

  1. Core Contribution and Differences from FDS (BeV2). Bi-FlowGS is the first to break down the barrier between diffusion-based methods and geometry-regularization-based methods, jointly enabling wide-range scene reconstruction and Gaussian geometry correction from sparse views.

    a. Unseen-View Expansion: Rendering improves over FDS across comparisons, with representative PSNR rising from 14.96 to 17.70 dB.

    b. Targeted Design of the Teacher Video Diffusion Model: G2V reduces AbsRel by 38.31% versus no flow conditioning (Table 5; flow visualizations: Figure 13).

    c. Effectiveness and Generality of Geometric Supervision: On GenFusion/GSFixer, V2G’s mean relative gains across eight geometry metrics are 17.39% / 41.42%, versus FDS’s 9.37% / 9.35%.

    d. Practical Application Value: Bi-FlowGS improves reconstruction across multiple datasets with wide-ranging viewpoints, demonstrating its practical value for sparse-view scene reconstruction.

  2. Analysis of Teacher-Flow Accuracy and Supervision Effectiveness (cHL7, BeV2).

    a. Reconstruction comparisons with alternative teacher flow: Restored-video supervision raises PSNR from 15.40 to 15.93 dB versus rendered flow; a separate restored/GT-video comparison yields 15.83 / 15.92 dB (GT-video comparison: Appendix E.5).

    b. Direct teacher-flow accuracy evaluation: Rendered/restored/GT-video flow yields EPE of 4.150 / 3.652 / 2.490 pixels (Appendix E.5; flow visualizations: Figure 18).

    c. Student-flow accuracy over training: Restored-video/rendered/no-flow supervision yields final student EPE of 3.4688 / 4.1726 / 4.4203 pixels (Appendix E.5; iteration visualizations: Figure 19).

  3. Experiments and Analysis of V2G’s Robustness to Erroneous Flow (WQuL, BeV2, cHL7).

    a. Filtering evaluation across restoration-quality levels: Errors decrease across groups; low-quality Fl-all falls from 50.65% to 32.18% (Appendix E.6).

    b. Extreme-failure filtering and failure-case analysis: Filtering rejects 80.64% of valid evaluation pixels; mean EPE is 8.141 pixels before filtering and 4.353 across retained pairs (Appendix E.6; failure visualizations: Figure 14).

    c. Reconstruction with and without masks under flow perturbations: Masked/unmasked/unperturbed PSNR is 15.87 / 15.68 / 15.93 dB, limiting degradation (Appendix E.6).

  4. G2V Conditioning-Prior Ablation (cHL7).

    Reconstruction comparison across G2V conditioning priors: Our conditioning/reprojection flow/Plücker yield AbsRel of 0.0182 / 0.0208 / 0.0244 (Table 5; flow visualizations: Figure 13).

  5. Controlled Comparison with Depth Supervision (BeV2). Flow/monocular depth/stereo depth/no geometric supervision yield Surface F-scores of 12.91% / 4.50% / 6.59% / 3.86% (Appendix E.4).

  6. Additional ArtiFixer Comparisons and Extended Applicability Validation (WQuL).

    a. Reconstruction comparison with ArtiFixer: Representative Bi-FlowGS/ArtiFixer PSNR is 19.33 / 19.29 dB.

    b. Plug-and-play evaluation on ArtiFixer: V2G reduces AbsRel from 0.1079 to 0.0467 and raises PSNR from 19.29 to 19.42 dB.

  7. Improved Manuscript Presentation (WQuL, cHL7). We simplified Figures 1–2, clarified implicit bidirectional co-refinement in Section 3 (Method), and reorganized Section 3.2 to align Equations (4)–(6) with the pipeline and clarify V2G/G2V’s complementary guidance.

Sincerely,

All Authors of Submission 3782

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.