acceptodds
Under review as a conference paper at ICLR 2027

ConFixGS: Learning to Fix Feedforward 3D Gaussian Splatting with Confidence-Aware Diffusion Priors in Driving Scenes

Abstract

Feedforward 3D Gaussian Splatting (3DGS) often struggles in trajectory-based sparse-view driving scenes. Existing Gaussian repair methods mainly target optimization-based 3DGS, while diffusion-based repair is typically restricted to iterative refinement near observed viewpoints, leaving feedforward 3DGS repair underexplored. We propose ConFixGS, a plug-and-play method that learns to fix feedforward 3DGS with confidence-aware diffusion priors. Starting from a pretrained feedforward model, ConFixGS generates diffusion-enhanced local pseudo-targets and validates them through reprojection-based cross-checking against support views. The resulting dense confidence maps guide refinement, enhancing reliable details while suppressing hallucinated or inconsistent evidence. On Waymo, nuScenes, and KITTI, ConFixGS improves challenging novel view synthesis, with PSNR gains of up to 3.68 dB and FID reduced by nearly half. Our results highlight confidence-aware fusion of generative priors and support-view consistency as a key principle for robust feedforward 3D driving scene reconstruction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D Gaussian Splatting, Feedforward Reconstruction, Diffusion Priors, Confidence-Aware Refinement, Sparse-View Reconstruction, Novel View Synthesis, Driving Scene Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sp5tOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a plug-and-play method for driving scenes 3DGS reconstruction, which uses diffusion prior to repair the novel view renderings and decide the reliability of novel view renderings by multi-view projection. The experiments are strong and convincing.

Critical strengths

This paper targets at the 3DGS reconstruction for driving scenes, which uses diffusion models to repair the novel view renderings. Different with most works, it decide the reliability of renderings by multi-view projection which greatly reduces the hallucination effects. The results are strong and convincing.

Critical weaknesses
  1. I think the contribution is over-claimed. At L051, "Recent diffusion-assisted 3DGS repair methods show that 2D generative priors can enhance degraded renderings and distill high-frequency details back into 3D representations. xxx. However, these methods are mostly designed for optimization-based 3DGS xxx". This kind of sentences make me fill that the author applied a completely different repair method that is specifically designed for feed-forward 3DGS. If my understanding is correct, the method firstly uses a feed-forward 3DGS model for initial 3DGS reconstruction, and use the similar way to repair the rendering with the additional confidence map, and finally use the novel view rendering with confidence map to optimize a 3DGS. It is still essential an optimization-based 3DGS method with refinement.
  2. The verification method of deciding the confidence is not novel. Please refer to [1][2].

[1] Empowering sparse-input neural radiance fields with dual-level semantic guidance from dense novel views [2] Geconerf: few-shot neural radiance fields via geometric consistency

Questions for authors
  1. Why did the author emphasize "local", like the statement "Although such local reconstructions are incomplete, they provide useful intermediate renderings that can serve as repair candidates for a global representation." If my understanding is correct, the repair candidates are sampled from nearby views of all input views.
Reviewer_4ZPyOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ConFixGS, a method for refining feedforward 3D Gaussian reconstructions of driving scenes. It generates pseudo-targets through local feedforward reconstruction and diffusion enhancement, then uses reprojection-based confidence to guide optimization of the global Gaussian model. ConFixGS+ additionally enhances the final rendered images. The method is evaluated on multiple driving datasets and feedforward backbones.

Critical strengths

The paper addresses a practical problem: improving feedforward reconstructions while limiting the influence of unreliable diffusion-generated supervision. The overall pipeline is clearly presented, and the experiments demonstrate consistent improvements over the initial reconstructions across multiple datasets and backbones. The supplementary runtime breakdown and robustness analyses provide useful insights into the method’s efficiency and behavior.

Critical weaknesses
  1. The SSIM results in Table 3 warrant further discussion. Difix3D achieves 0.7817, compared with 0.7193 for ConFixGS. With final postprocessing, Difix3D+ also achieves a slightly higher SSIM than ConFixGS+ (0.6856 vs. 0.6808). Although ConFixGS improves other metrics, the paper does not directly explain this trade-off. Given the presence of moving objects in the driving sequences, it would also be helpful to explain where the baseline using a static 3DGS representation performs well and whether its SSIM advantage primarily comes from static regions.
  2. The distinction between feedforward repair and optimization-based repair is not sufficiently established. The confidence estimation and Gaussian refinement stages appear applicable to optimization-based reconstructions as well. Replacing local feedforward inference with direct rendering from an existing optimized Gaussian model seems to produce a similar repair pipeline. While this broader applicability could be valuable, the current discussion does not clearly identify which aspects of feedforward reconstruction require a distinct repair strategy.
  3. Figure 4 provides examples in which local feedforward inference produces better renderings than global inference, but the consistency of this advantage across the datasets remains unclear. Scene-level or dataset-level statistics would provide stronger support for this observation. An ablation that applies diffusion directly to renders from G0G_0, while keeping confidence estimation and subsequent optimization unchanged, would also help establish the contribution of the local feedforward branch.
Questions for authors
  1. In the “w/o Diffusion” ablation, the targets are described as coming from “local or G0G_0 renderings.” Which source is used in the reported results? If both are used, how do you choose between them?
  2. How is the rendering from G0G_0 incorporated into local feedforward inference? Is it passed directly to the backbone as an additional input image, or is any preprocessing or special conditioning applied? How many real support images are included, and how are they selected?
  3. What explains the SSIM difference in Table 3? Does the advantage of Difix3D mainly come from static regions, or does it also hold for moving objects?
Reviewer_2NdZOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes ConFixGS which is an post-optimization module of the feedforward 3DGS models to improve the rendering quality in the static driving scenarios. The uses frozen feedforward 3DGS model to initialize the 3D Gaussian from sparse video input and use frozen Diffusion model to provide the pseudo groundtruth. After confidence-aware per-scene optimization, the repaired 3DGS can improve the rendering quality at the novel views. The evaluation on driving benchmarks demonstrate the effectiveness of the repairing design.

Critical strengths
  1. The method is model-agnostic for 3D GS repairing. The paper validate the design on different backbones. The method compare with feedforward and per-scene optimization method and achieve better performance.
  2. The training-free confidence weighting method is intuitive and the ablation study shows the importance of this design.
Critical weaknesses
  1. In support-validated confidence design, the paper applies the spatial smoothing. However, the heuristic smoother may incorrectly blur the confidence weight and further cause the blurring in the rendering, especially on the sharp boundary where depth changes abruptly or the spatial frequency is high. These problem may be not clear nearby the input views, but when the novel view shifts largely to the side or top of the targets, the problem is likely to be obvious. If this is not an actual problem, the authors are expected to show demos with larger view shifts.
  2. This model uses DepthSplat and WorldMirror as backbone, however, they are not dedicated model for driving scenario so the proposed method is easier to demonstrate the improvement on driving cases. The pixelSplat and MVSplat have the similar issues. And the feedforward-only method AnySplat achieves clearly better LPIPS than ConFixGS. As a result, the experiments can not strongly validate this proposed method.
  3. The model aims to improve rendering quality in the driving domain where dynamics play a crucial role. Unfortunately, the backbone is designed for static reconstruction, and the proposed method has a static underlying assumption such as the operations of unprojection and reprojection, so the model inherently can not solve the dynamics. As Table 8 and 9 shows, the dynamic region has much smaller improvement than the static region. Furthermore, the video demos seems to only compare methods on the static scenes. As a result, the fixing of this method is limited to static background in the driving scenes.
Questions for authors

Besides of questions in Weakness section, there is one additional question:

In table 8 and 9, only the PSNR is reported. Could the authors also report the SSIM, LPIPS, and qualitative results? The small improvement on dynamic PSNR only can not strongly validate the better quality on the dynamic regions.

Comments on the paper

1 comment

Dear Area Chair and Reviewers,

We sincerely thank the Area Chair for coordinating the review process and the reviewers for their constructive suggestions and insightful questions. Your feedback has been valuable in improving our paper.

As the rebuttal period comes to an end, we have uploaded a revised paper with the additional results and clarifications from the reviews and discussion. All changes are marked in blue.

We appreciate the follow-up feedback received so far and welcome any further comments or suggestions. We would be happy to provide additional clarifications or revise the paper further.

Sincerely,

The Authors of Paper 55052

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.