acceptodds
Withdrawn from ICLR 2027

GeoFix: Geometry Fixing for 4D Consistent Generation

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

32 views

Abstract

Video diffusion models increasingly serve as priors for 4D generation, but the generated videos often lack geometric consistency. Recent approaches address this through post-training by designing geometric rewards, yet this can degrade texture and fine detail, requiring auxiliary appearance rewards and careful reward balancing. To address this challenge, we propose GeoFix, a unified framework that improves geometric consistency and visual quality with a single geometric reward, applicable to both training and inference-time reward alignment. Our reward measures cross-view depth consistency in static scene regions, targeting structural inconsistencies without directly constraining appearance or penalizing valid object motion. Crucially, depth agreement increases with camera field-of-view overlap even in real videos, allowing a smaller camera movement to be mistaken for better geometry. We account for this effect by calibrating the reward against the overlap-dependent trend. GeoFix achieves a improvement in WorldScore 3D consistency and a improvement in general video quality measured by VBench.

Primary area:
generative models
Keywords:
Video Diffusion Model, Generative Models, 4D Gaussian Splatting

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_TXNtOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoFix improves geometric consistency through FOCAL, a reward measuring static-region depth agreement and correcting for camera field-of-view overlap. A stitched reconstructor computes geometry directly from video latents. The framework supports post-training and inference-time alignment, with experiments reporting improved consistency and quality of generated 4D scenes.

Critical strengths
  1. The observation that geometric rewards can favor smaller camera movements is useful. Calibrating depth consistency against viewing conditions is a well-motivated contribution beyond existing latent-space geometric alignment, such as VGGRPO.

  2. The controlled training comparisons are encouraging: methods share the SFT data, training budget, and stitched reconstructor. Table 1 shows improvements in both geometric consistency and rendered-video quality, while the corruption experiments provide additional evidence that the reward responds to structural errors.

Critical weaknesses
  1. The calibration does not affect the reported DPS guidance. Appendix A.1.5 states that the FOV term is non-differentiable and treated as a constant offset. Consequently, the implemented gradient of FOCAL equals that of the uncalibrated reward. The DPS results therefore support depth-based guidance, but do not establish the benefit of the proposed calibration in gradient-based alignment.

  2. The calibration’s ability to prevent motion suppression is insufficiently validated. Table 4 reports overall quality and dynamic degree, but neither direct camera-motion measurements nor geometric-consistency scores. Dynamic degree alone cannot distinguish camera motion from object motion or artifacts. Moreover, Table 1 reports lower motion magnitude for GeoFix (35.44) than the base model (45.17). This does not prove motion collapse, but warrants a more direct analysis. The calibration dataset, fitting protocol, and held-out statistics also need clearer specification.

Questions for authors
  1. Given the constant-offset implementation, should DPS produce identical results with and without FOV calibration under otherwise identical settings?

  2. Can Table 4 include geometric consistency and separate camera-motion and object-motion measurements?

  3. Which videos and splits were used to estimate the calibration coefficient, and how stable is it across scene types and camera trajectories?

Reviewer_HG4pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This manuscript introduces a geometric reward that can be used to improve geometric consistency in both post-training and inference. The reward is basically cross-view depth-consistency in static regions, with the calibration for reducing the FOV overlap bias. Depth and camera parameters are predicted by stitching NeoVerse to the output latent, making it also compatible with gradient-based guidance.

Critical strengths
  1. This paper convincingly points out and verifies the bias in the depth consistency-based reward toward larger FOV overlap, while providing a reasonable explanation and analysis in the appendix.
Critical weaknesses
  1. Although the phenomenon about overlap bias is demonstrated relatively sufficiently, the proposed solution, i.e., the calibration mechanism, receives much less validation. It is suggested to provide more qualitative ablations about it beyond two numbers in Table 4. In addition, this design relies on several heuristic components.
  2. The role of the proposed calibration in DPS-style guidance is unclear, since the magnitude of camera motion is typically determined in the early denoising process when accurate depth predictions may still be difficult.
Questions for authors
  1. Given that the stitched reconstructor is trained on clean latent z0z_0, does applying it to z0∣tz_{0|t} in DPS-style guidance introduce a domain gap? Could incorporating some in-domain samples during training improve its performance?
  2. Computing FOV overlap using only the first and last frames seems somewhat heuristic. In certain scenes, it may not be a good metric for the magnitude of camera movement, e.g., a 360° orbit video.
Reviewer_5nL6OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents GeoFix, a framework for improving the geometric consistency of video diffusion models for 4D generation. The method introduces FOCAL, a geometric reward based on cross-view depth reprojection consistency in predicted static regions. The authors observe that raw depth-consistency scores are biased by camera field-of-view overlap: videos with smaller camera motion tend to receive higher scores even when their geometry is not necessarily better. FOCAL attempts to remove this bias by calibrating the raw score against an overlap-dependent trend estimated from real videos. Experiments on video generation and 4D scene generation report improvements on WorldScore, VBench, and camera-accuracy metrics.

Critical strengths
  1. The paper identifies a meaningful bias in existing geometric rewards. Raw cross-view depth-consistency scores can favor videos with greater field-of-view overlap and therefore smaller camera motion. Recognizing this failure mode is relevant to reward-based alignment for video and 4D generation.

  2. FOCAL provides a simple mechanism for reducing this bias. The overlap-dependent calibration is conceptually intuitive and can be applied to a geometric reward without introducing an additional appearance reward or requiring the balancing of multiple reward terms.

  3. The reward focuses on static scene regions. Excluding predicted dynamic regions from depth reprojection prevents valid object motion from being directly penalized as a geometric inconsistency.

Critical weaknesses
  1. The method is technically reasonable, but its main contribution appears incremental over VIST3A-style latent stitching and reward alignment. The primary novelty is an overlap-dependent scalar correction to an existing depth-consistency reward. Moreover, this correction does not affect the DPS gradient because the FOV term is treated as non-differentiable.

  2. With the authors’ current visualization style, including the supplementary videos, I cannot clearly identify where the improvements are. The authors could use bounding boxes or annotations to highlight regions with improved geometric quality, making the claimed improvements easier for readers to perceive.

  3. The reward construction and evaluation are heavily coupled through the same stitched 4D reconstructor. This reconstructor is used to compute the reward, produce the final 4D representation, render videos for evaluation, and estimate camera trajectories for the camera-accuracy metrics. Consequently, the reported gains may partly reflect improved compatibility with this particular reconstructor rather than a general improvement in geometric consistency. Evaluation with an independent reconstructor or ground-truth geometry would strengthen the claim.

  4. The main ablation for FOV calibration is insufficient. Table 4 reports only the overall VBench score and dynamic degree. These results suggest that calibration prevents motion suppression, but they do not demonstrate that it improves geometric correctness. Reporting 3D consistency, camera errors, and independent geometric metrics for different values of gamma would be more informative.

  5. The calibration assumes a single global log-linear relationship between FOV overlap and depth consistency. However, this relationship may also depend on scene depth, occlusion, focal length, motion type, dynamic-region ratio, and reconstruction uncertainty. The paper does not provide sufficient evidence that one fixed calibration coefficient generalizes across datasets and camera trajectories.

Questions for authors

Please refer to the weaknesses.

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

Related papers

Discussion (0)

Sign in to comment.