acceptodds
Withdrawn from ICLR 2027

ViewRobust-3D: Benchmarking and Improving Viewpoint Robustness in Single-View 3D Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In recent years, single-view 3D generation has advanced rapidly and shown great potential in many areas. However, existing models are primarily developed and evaluated under canonical views, leaving their robustness to broader view variations poorly understood. We observe that non-canonical inputs (e.g., high elevation or non-zero roll angles) can lead to geometric distortions and structural errors. Motivated by this observation, we introduce ViewRobust-3D, a controlled benchmark to systematically analyze the impact of view variations on single-view 3D generation. By isolating view factors through controlled changes in azimuth, elevation, and roll, the benchmark reveals performance degradation and interactions between these factors across multiple models. Guided by these findings, we further propose ViewNorm, a lightweight view normalization strategy to normalize non-canonical inputs toward canonical views. Experiments validate ViewRobust-3D in exposing robustness gaps and demonstrate the effectiveness of ViewNorm in improving the performance of single-view 3D generation models under non-canonical views. Our benchmark and method will be publicly released upon publication.

Primary area:
generative models
Keywords:
single-image-3d-generation

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_ewQkOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ViewRobust-3D, a controlled benchmark for studying viewpoint robustness in single-view 3D generation. The benchmark evaluates several representative models under controlled changes in azimuth, elevation, and roll, and reports that current methods degrade substantially under non-canonical viewpoints, particularly for non-zero roll and high elevation.

The paper further analyzes the interaction between elevation and roll and proposes CanoView, a plug-and-play input canonicalization pipeline. CanoView estimates and corrects roll through image rotation and transforms high-elevation inputs toward a canonical elevation using a fine-tuned Stable Zero123 model before applying the downstream 3D generator. Experiments show that this preprocessing improves reconstruction quality for several existing models under non-canonical views.

Critical strengths
  1. The paper studies an important and underexplored robustness issue. Existing single-view 3D generation methods are commonly demonstrated on canonical inputs, and the controlled evaluation provides useful evidence that elevation and roll can significantly affect reconstruction quality.

  2. The benchmark design enables controlled analysis of viewpoint factors across multiple representative 3D generation architectures. The elevation-roll interaction analysis is also more informative than simply reporting average degradation under arbitrary viewpoints.

Critical weaknesses
  1. The benchmark is based primarily on controlled synthetic renders from Objaverse objects, while the motivation emphasizes real applications such as robotics and user-captured images. Such real inputs include clutter, occlusion, truncation, backgrounds, lighting variation, and camera differences that are not represented by the current controlled setup. Without evaluation on real non-canonical photographs, it is unclear how well the reported robustness trends generalize to practical inputs.

  2. The main geometric metric does not clearly disentangle orientation errors from shape reconstruction errors. Predictions are aligned only in azimuth, while elevation and roll are intentionally left unaligned. As a result, Chamfer Distance penalizes both incorrect object geometry and a globally misoriented but otherwise geometrically correct reconstruction. Since the paper aims to characterize viewpoint-induced 3D reconstruction degradation, separating global orientation error from orientation-normalized geometry error would make the conclusions substantially clearer.

Questions for authors
  1. Do the authors have results on real non-canonical images, such as handheld object photographs or robotics-camera inputs? If so, do the same elevation and roll robustness trends observed on controlled renders persist?

  2. Can the authors report geometry metrics after full rigid orientation alignment, together with a separate orientation-error metric? This would help determine whether non-canonical inputs primarily affect shape reconstruction or global coordinate orientation.

  3. How much of CanoView’s improvement comes simply from moving the input distribution closer to canonical training views? In particular, how does the method compare with simpler canonicalization baselines that do not use a diffusion-based novel-view synthesis model?

  4. Since elevation correction relies on Stable Zero123, how sensitive is downstream reconstruction to hallucinations or view-synthesis errors introduced by this additional generative stage?

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

This paper introduces ViewRobust-3D, a benchmark for systematically evaluating viewpoint robustness in single-view 3D generation. It examines the effects of azimuth, elevation, and roll on four representative 3D generation models using controlled renderings from Objaverse. The results reveal substantial performance degradation under non-canonical viewpoints, particularly for elevation and roll variations, as well as non-additive interactions between these factors. Based on these observations, the authors further propose CanoView, a plug-and-play preprocessing framework that estimates input viewpoints and transforms non-canonical images toward canonical views without retraining downstream 3D generation models. Experiments demonstrate improved reconstruction quality across most evaluated models and viewpoint settings.

Critical strengths

1. Limited disentanglement of geometric and orientation errors.

The benchmark evaluates Chamfer Distance after azimuth-only alignment, intentionally preserving elevation and roll discrepancies. While this is reasonable for assessing overall viewpoint robustness, the metric conflates geometric reconstruction errors with global orientation errors. Consequently, the relative contributions of these two failure modes remain unclear. Reporting fully pose-aligned CD alongside the current metric would strengthen the interpretation of the benchmark results.

2. Insufficient characterization of benchmark representativeness and statistical reliability.

The benchmark contains approximately 350 objects from an aesthetically filtered Objaverse subset. However, the paper provides limited information about category diversity, geometric complexity, and object symmetry. Additionally, the confidence interval estimation in Table 2 is not sufficiently described, particularly regarding dependencies between multiple observations of the same object. These omissions make it difficult to assess the generalizability and statistical reliability of the findings.

Critical weaknesses
  1. Could the authors additionally report fully pose-aligned Chamfer Distance to distinguish geometric reconstruction errors from orientation errors? In particular, do the elevation–roll interaction patterns persist under this evaluation?

  2. Could the authors provide more details about the distribution of object categories, geometric complexity, and symmetry in the benchmark to clarify its representativeness?

  3. How are the 95% confidence intervals in Table 2 computed? Do they account for correlations among different viewpoints of the same object?

Questions for authors

1. Evaluation protocol. Could the authors additionally report Chamfer Distance after full rotational alignment, alongside the current azimuth-aligned results? This would help distinguish geometric reconstruction failures from global orientation errors and clarify whether the observed viewpoint sensitivity persists independently of orientation discrepancies.

2. Benchmark representativeness and statistical validity. Could the authors provide statistics on object categories, geometric complexity, and symmetry in the benchmark? Additionally, how are the 95% confidence intervals in Table 2 computed, and do they account for correlations among different views of the same object?

3. Interpretation of the elevation–roll interaction. The negative interaction effects in Table 2 are calculated in log-CD space. Could the authors clarify whether similar interaction patterns remain when using absolute CD differences and fully pose-aligned geometric errors? This would strengthen the interpretation that high elevation genuinely attenuates roll sensitivity rather than merely reducing its relative effect due to higher baseline errors.

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

This paper investigates the robustness of single-view 3D generation models to non-canonical viewpoints. The authors introduce ViewRobust-3D, a controlled benchmark that evaluates the effects of azimuth, elevation, and roll on four representative 3D generation models. The experimental results show substantial performance degradation under non-canonical viewpoints, particularly for roll variations, and reveal a non-additive interaction between elevation and roll.

To mitigate these issues, the authors propose CanoView, a preprocessing framework that estimates and normalizes the input viewpoint before passing the image to existing 3D generation models. CanoView combines roll correction with diffusion-based elevation transformation. Experiments demonstrate improvements in Chamfer Distance and PSNR across most evaluated models and viewpoint conditions.

Critical strengths
  1. Systematic evaluation of viewpoint robustness.

The paper addresses a relevant limitation of existing single-view 3D generation models, which are often evaluated primarily under canonical viewpoints. The controlled benchmark provides useful empirical evidence of performance degradation under elevation and roll variations. In particular, the observed non-additive interaction between these factors is an interesting finding that extends beyond simply reporting performance drops under challenging viewpoints.

  1. Practical improvement without retraining downstream models.

CanoView provides a model-agnostic preprocessing strategy that can be applied to existing 3D generation models. The improvements across multiple architectures, particularly the 33.9% CD reduction for TRELLIS on the Joint-Noncanonical subset, demonstrate the practical potential of input viewpoint normalization. The ablation study also provides evidence for the complementary contributions of the proposed components.

Critical weaknesses
  1. Insufficient separation of geometric reconstruction errors and orientation errors.

A major concern is the evaluation protocol used to support the paper's conclusions about viewpoint robustness. As described in Section 3.1, predicted and ground-truth 3D objects are aligned only along the azimuth direction, while elevation and roll discrepancies are preserved.

Although this protocol is reasonable for evaluating orientation-sensitive reconstruction, the reported Chamfer Distance consequently reflects both geometric discrepancies and global orientation errors. For instance, a geometrically accurate reconstruction with an incorrect global orientation could produce a large CD.

This distinction is particularly important because the paper attributes performance degradation to viewpoint-induced reconstruction failures and further analyzes the interaction between elevation and roll. Without separately reporting orientation errors and rotation-invariant geometric reconstruction quality, it is difficult to determine how much of the observed degradation is attributable to actual geometric errors. It is also unclear whether the reported elevation-roll interaction persists when global orientation differences are removed.

I believe that disentangling these two sources of error is important for substantiating the paper's central empirical claims.

  1. Limited methodological novelty and insufficient comparisons with straightforward viewpoint normalization baselines.

The proposed CanoView framework primarily combines existing components, including DINOv3-based angle estimation, 2D image rotation, a fine-tuned Stable Zero123 model, and MeanShift postprocessing. While this combination is practically motivated, the paper does not clearly establish what methodological contribution distinguishes it from a straightforward pose estimation and novel-view synthesis pipeline.

More importantly, Table 3 compares CanoView only against direct 3D generation without viewpoint normalization. Although the component ablations in Table 4 are useful, they do not establish whether the proposed design outperforms simpler alternatives, such as viewpoint correction using ground-truth camera angles or an off-the-shelf novel-view synthesis model.

Consequently, it remains unclear whether the improvements arise from the specific design of CanoView or primarily from the general benefit of converting non-canonical inputs into canonical views.

Given the relatively limited algorithmic novelty, stronger comparative evidence is needed to establish the significance of the proposed approach.

Questions for authors

Q1. Disentangling geometry and orientation errors.

Could the authors report a rotation-invariant geometric reconstruction metric, such as CD after full rigid rotational alignment, alongside a separate orientation error metric? In particular, does the negative elevation-roll interaction reported in Section 3.4 remain significant when orientation discrepancies are removed? This analysis would help clarify whether the observed interaction reflects geometric reconstruction behavior or is influenced by the orientation-sensitive evaluation protocol.

Q2. Validating the contribution of CanoView.

Could the authors compare CanoView against simpler viewpoint normalization baselines, such as using ground-truth camera angles for normalization and using a pretrained Stable Zero123 model without task-specific fine-tuning? These comparisons would help establish the additional benefit of the proposed pipeline over general viewpoint canonicalization. Reporting the additional inference cost would also help assess its practical utility.

Reviewer_p53rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary
  • The paper studies how the input viewpoint—azimuth, elevation, and roll—affects single-view 3D generation.
  • The benchmark contains approximately 350 Objaverse objects, organized into 588 object–azimuth groups. Each group includes four conditions: canonical view (C), high elevation (E), random roll (R), and combined elevation and roll (ER), giving 2,352 inputs overall. The authors evaluate TripoSR, SPAR3D, TRELLIS.2, and Hunyuan3D 2.1 using Chamfer distance after azimuth-only ICP alignment and four-view PSNR.
  • All four models perform worse outside the canonical setting, with roll generally causing larger errors than elevation. A difference-in-differences analysis of log-CD finds a negative elevation–roll interaction for every model.
  • The paper also introduces CanoView, a preprocessing pipeline that requires no retraining of the generators. A DINOv3-based estimator corrects roll when its magnitude is at least 45 degrees. Inputs with estimated elevation of at least 50 degrees are passed to a fine-tuned Stable Zero123 model to produce a view at 35 degrees, followed by MeanShift filtering. CanoView reduces CD by 12–34% on the R and ER subsets, while its results on E are mixed.
Critical strengths
  • Evaluation on canonical views can give an incomplete picture of how image-to-3D models behave on casually captured images, and a controlled study of viewpoint sensitivity is useful.
  • CanoView is simple to apply across different generators.
Critical weaknesses
  • The evaluation does not clearly separate orientation error from shape error. This is my main concern. Predictions are aligned to the ground truth only in azimuth. If a model reconstructs an object in the input camera’s frame, a rolled or elevated input can produce a geometrically accurate object that nevertheless receives a high CD because its orientation differs from the reference. The pineapple and cat examples in Figures 1 and 7 appear consistent with this possibility.
  • The CanoView experiments leave several basic questions unanswered. For roll correction, the benchmark applies an in-plane image rotation, and the estimator is trained to recover that same transformation. This is a useful controlled test, but it provides limited evidence about robustness beyond the benchmark’s transformation setup. The related-work discussion also misses relevant approaches. One-2-3-45 (NeurIPS’23) and Era3D (NeurIPS’24) already estimate or handle input elevation; Orient Anything (ICML’25) estimates object orientation from a single image; and CUPID (arXiv:2510.20776) jointly generates shape and camera pose. These methods deserve discussion and, where appropriate, comparison. Finally, the benchmark, estimators, and ERN all use Objaverse. There is also a plausible risk of overlap with the generators’ training data: TRELLIS-500K, for example, draws on aesthetic-filtered Objaverse-XL data from Sketchfab. The paper does not establish the extent of this overlap or evaluate on real photographs or an independently held-out asset collection. This limits how broadly the conclusions can be drawn.
Questions for authors
  • Can you repeat Tables 1 and 3 using full rotational alignment, or similarity alignment where appropriate, and report orientation error separately? How much of the viewpoint degradation and CanoView improvement remains after orientation is accounted for?
  • How does CanoView compare with unmodified Stable Zero123, upright correction based on Orient Anything, and a simple test-time strategy that tries rotations in multiples of 90°?
  • Can you assess likely overlap between the benchmark assets and the generators’ training data?

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

Related papers

Discussion (0)

Sign in to comment.