acceptodds
Withdrawn from ICLR 2027

AlphaRefine: Opacity-Routed Refinement for Feed-Forward Driving Scene Reconstruction

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Feed-forward driving scene reconstruction has emerged as a scalable alternative to per-scene optimization, but still struggles to produce globally consistent appearance and fine structural details. Recent efforts close this gap through diffusion-based post-hoc refinement, which applies a uniform and computationally heavy generative enhancement. However, this uniform treatment overlooks the actual error structure of feed-forward renderings. Through empirical analysis on multiple feed-forward models, we observe that rendering errors are not spatially homogeneous but strongly correlated with accumulated opacity, where low-opacity regions are dominated by low-frequency appearance shifts and high-opacity regions suffer primarily from high-frequency structural loss. Building on this observation, we identify a complementary correspondence between these two error regimes and two in-context correction signals encoded in the source view. Specifically, the source image serves as appearance context for low-frequency correction, while the source-view residual provides error context for high-frequency recovery. Based on this insight, we introduce **AlphaRefine**, a lightweight post-hoc refinement framework that routes the two corrections by accumulated opacity, applicable to any frozen feed-forward driving reconstruction model. Under the same evaluation protocols, our method achieves new state-of-the-art among feed-forward methods on both Waymo and nuScenes, boosting STORM to dB ( dB) and DrivingForward to dB ( dB), while running over faster than single-step diffusion refiners at only ms per image. Code will be released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
3D Reconstruction, Gaussian Splatting

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

The paper suggests a post-hoc fix mechanism for feed-forward driving scene reconstruction. Similar to recent methods that use diffusion to fix reconstruction artifacts, the suggested method uses a dedicated small model to fix both low and high frequency errors, using a learned opacity-conditioned model and projection-based sampling. The suggested method shows generalization across different feed-forward models, is faster than alternatives, and has good quantitative performance.

Critical strengths
  • The paper is mostly well-written and easy to follow.
  • The suggested method looks practical for driving-scene reconstruction.
  • The authors are willing to publish their code for future research.
Critical weaknesses
  • The papers mention the insight and analysis of the correspondence between predicted opacity and the frequency of the errors. However, it briefly mentions it in the introduction, and then states this as a fact instead of proving this claim. While there are more analysis details in the supplementary material, this analysis is still not convincing enough to be claimed as a main contribution. Moreover, while the paper emphasizes this insight, it's not clear how this fact affected the design choices, as there is no explicit frequency-based fixing.

  • Limited comparisons - The method claims superiority over diffusion-based methods, but it seems that it will only work when the rendered view is very close to a reference view, such that the initial rendered target view is good enough and that the projection correction makes sense. However, most diffusion-based methods operate in a very sparse setting where the generative priors allow for reconstruction with very limited information. This comparison setting should be made clearer, and the model should be evaluated under different sparsity configurations. Under these sparse settings, FID (which is generative in nature) should also be reported.

  • Limited novelty - While effective in the evaluated setting, post-fixing feed-forward reconstruction methods are not new. As Artifixer already uses opacity as a measure of how much "fixing" is needed, this method seems to be mostly better engineering for the setting than a novelty.

Questions for authors
  • It's not clear how the DCN module differentiates rendering errors from errors caused by the dynamic regions.
  • The qualitative figures are extremely low quality, making it very hard to compare the methods.
Reviewer_kcUoOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies post-hoc refinement for feed-forward driving scene reconstruction. It proposes AlphaRefine, a lightweight two-branch refinement framework (CAVC\CVRC ) that uses source-view appearance context for low-frequency correction and transported source residuals for high-frequency recovery, routed by opacity. The method is evaluated on Waymo and nuScenes.

Critical strengths
  1. The paper is clearly written and easy to follow.
  2. The experimental results are promising. As a lightweight plug-and-play post-refinement module, AlphaRefine can be integrated into existing feed-forward reconstruction models (e.g., STORM and DGGT), achieving a substantial PSNR improvement of up to 2.51 dB, as shown in Table 1.
Critical weaknesses
  1. Imprecise description of prior work. In Line 125, the statement regarding per-scene optimization following the emergence of 3DGS is potentially misleading, as earlier NeRF-based methods also relied on per-scene optimization.

  2. Limited discussion of related work. The Related Work section mainly lists existing methods without sufficiently analyzing their technical differences and connections to the proposed approach. The wiriting style somehow like AI-Generated contents.

  3. Limited improvement to the underlying 3D representation. AlphaRefine performs post-hoc refinement entirely in image space without improving the underlying 3D Gaussian representation. This limits its practical value for applications requiring consistent 3D scene reconstruction.

  4. The central motivation assumes a strong relationship between accumulated opacity and rendering error characteristics. However, the current empirical analysis does not sufficiently establish this relationship or demonstrate that opacity is the most appropriate routing signal.

  5. The paper claims superior rendering quality over diffusion-based refinement methods. However, Table 2 does not consistently support this claim. e.g. DGGT + Difix3D+ achieve the best results.

Questions for authors

no

Reviewer_B7LdOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

AlphaRefine is a lightweight framework for refining 2D renderings from frozen feed-forward driving scene reconstruction models. Its central contribution is an empirical association between accumulated opacity and rendering error types, which motivates complementary corrections: CVAC uses source images to correct color and luminance in low-opacity regions; CVRC geometrically transfers source-view rendering residuals to restore textures and edges in high-opacity regions; and DCN learns sampling offsets to mitigate motion-induced misalignment. Experiments cover STORM, DGGT, and DrivingForward on Waymo and nuScenes. Trained only on Waymo, the refinement modules improve PSNR by 2.46–3.27 dB across all six backbone–dataset combinations, with consistent SSIM and LPIPS improvements. Ablations support branch complementarity and dynamic correction. The framework has approximately 3M parameters and reports 6 ms per image at 160×240 resolution, roughly 9× faster than single-step Difix3D+.

Critical strengths
  1. Substantial gains across backbones and datasets. All six backbone–dataset combinations in Table 1 improve on all three metrics, with substantial PSNR gains of 2.46–3.27 dB. Although the refinement modules are trained only on Waymo, they consistently improve their respective baselines on nuScenes, indicating useful transferability. Tables 3 and 8, together with additional visualizations, further suggest that the gains are not confined to one architecture or a few displayed examples. 2.A compact design makes effective use of source observations Appearance correction and geometric residual transfer form a simple, motivated combination, and the full model outperforms single-branch variants in Tables 3 and 8. DCN improves dynamic-region PSNR by 0.30, 0.25, and 0.51 dB for STORM, DGGT, and DrivingForward, respectively, demonstrating measurable value from motion correction.
Critical weaknesses
  1. The independent benefit of opacity routing is not established. A central contribution is routing complementary corrections using opacity. However, the experiments primarily establish a correlation between opacity and error characteristics and demonstrate the effectiveness of combining two branches; they do not isolate the benefit of routing through controlled comparisons. Under the described method, the branches evaluated in Tables 5, 9, and 10 remain opacity-gated, so their regional specialization may partly reflect the imposed spatial weighting. These results are consistent with the design motivation but do not establish an advantage over alternative fusion strategies. Without matched controls that retain both branches while removing spatial gating or using constant fusion weights, it is difficult to assess the additional benefit attributable to the central technical contribution.
  2. Broad SOTA claims are insufficiently supported across protocols and resolutions. Appendix B performs refinement and metric computation at 160*240, including resized higher-resolution DrivingForward outputs and ground truth. This supports controlled comparisons at the shared low resolution but does not establish superior detail recovery at native resolution. Moreover, the three backbones use different camera configurations or temporal sampling protocols, so values across their rows do not establish an overall ranking under a unified benchmark. Although improvements within each backbone remain compelling, the broad SOTA claims in Section 4.2 require a clearer connection to competing methods evaluated under matched protocols. This limits my assessment of the claimed contribution’s scope, particularly its ability to recover high-frequency details.
Questions for authors

1.What is the precise scope of the SOTA claim? Which competing methods and shared protocols underpin the SOTA claims associated with 28.32 dB on Waymo and 29.83 dB on nuScenes in Section 4.2? Please clarify whether the comparisons match input cameras, context frames, target frames, and evaluation resolution, and whether final metrics exclude input context frames. If SOTA refers only to the particular backbone–refiner combinations evaluated in this paper, could the claim be scoped accordingly? 2. Do detail-recovery gains depend on low-resolution evaluation? For a backbone whose native resolution exceeds 160240 such as DrivingForward, is there evidence that AlphaRefine still outperforms the unrefined baseline and the principal competing refiner at native resolution? If such experiments are available, please report PSNR, SSIM, and LPIPS under matched conditions. Otherwise, please clarify that the detail-recovery claim has currently been validated only at 160240.

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

Related papers

Discussion (0)

Sign in to comment.