- Rating
- 2
- Confidence
- 4
- Informativeness
- 1
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper presents a method for reconstructing continuous implicit fields from multi-view images without requiring known camera poses, using a VGGT-like framework. The core idea is straightforward. The method is trained end-to-end, similar to other feed-forward reconstruction approaches, but leverages the VGGT architecture to enable pose-free reconstruction through alternating frame-wise and global self-attention. The proposed method extends this architecture with a volume rendering head that samples view features from the VGGT latents and predicts SDF and color values to render images, similar to existing volume rendering approaches. The renderer is supervised using pixel-space losses, and query points for volume rendering are sampled using the predicted camera poses. The method employs an existing volume rendering approach, VolSDF, for surface reconstruction. Results in the paper shows that the method is significantly better than generalizable surface reconstruction approaches and performs at-par or little bit inferior that per-scene optimization methods.
Critical strengths
- The paper is well-written and easy to follow, with clear explanations of the methodology and implementation details.
- The method demonstrates significant improvements in reconstruction quality over previous generalizable surface reconstruction methods.
Critical weaknesses
-
Limited Technical Novelty: Despite being well-written, the paper offers limited technical novelty, as most components of the proposed pipeline are based on existing approaches. Similar to how VGGT employs different prediction heads for camera poses, depth, and point trajectories, this method introduces an additional volume rendering head for surface reconstruction, which appears to be a relatively straightforward extension of the architecture. Furthermore, the other components required for volume rendering are already well-established in the literature, such as feature sampling in PixelNeRF and surface reconstruction using VolSDF. Therefore, I find the overall technical contribution of the paper to be limited.
-
Improvements Over Per-Scene Optimization Methods: Although the method achieves significant improvements over previous generalizable reconstruction approaches, its improvements over per-scene optimization methods are marginal or, in several cases, worse in terms of metrics such as precision, accuracy (Acc), and F-score. Additionally, the paper does not appear to compare against sufficiently fast per-scene optimization baselines. For example, MonoSDF uses an MLP-based architecture, which is generally slow to optimize. However, MonoSDF also has an Instant-NGP variant that could potentially converge much faster than the 22 hours reported in the paper. Therefore, I believe the evaluation would benefit from comparisons with more efficient per-scene optimization baselines.
-
Unclear Contribution of the Proposed Method: It is difficult to determine whether the improvements over previous generalizable reconstruction methods stem from the proposed approach itself or primarily from the stronger pretrained VGGT backbone. The paper does not clearly isolate the contribution of the proposed components from the benefits of using a more powerful backbone architecture.
Questions for authors
-
I would like to see a more fair comparison with fast per-scene optimization baselines like instant-ngp (neuralangelo), 2D gaussian splatting and its varient.
-
I would like to hear from the authors about what is the novelty of their method.
5 comments
Author Response - Part 2
W3 - Isolate the contribution of the proposed method
Thanks for the suggestion. To isolate the contribution of our proposed method, we reevaluate previous generalizable baselines using the poses and geometry predicted by VGGT, as shown in Table (b).
Table (b). Comparison with generalizable baselines adapted to use VGGT-predicted poses and geometry.
| Method | Acc ↓ | Comp ↓ | Chamfer distance ↓ | Prec ↑ | Recall ↑ | F-score ↑ |
|---|---|---|---|---|---|---|
| SimpleRecon+VGGT | 0.061 | 0.076 | 0.068 | 0.532 | 0.529 | 0.529 |
| FineRecon+VGGT | 0.094 | 0.053 | 0.074 | 0.505 | 0.616 | 0.552 |
| MeshSplat+VGGT | 0.094 | 0.142 | 0.118 | 0.289 | 0.329 | 0.306 |
| TriSplat | 0.084 | 0.121 | 0.103 | 0.389 | 0.543 | 0.453 |
| IVGT(Ours) | 0.039 | 0.045 | 0.042 | 0.756 | 0.694 | 0.723 |
We retain TriSplat’s original results because it already uses Pi3 [4], another VGGT-initialized geometry model. We think these results provide further evidence for the contribution of our proposed method.
[4] Wang, Y., Zhou, J., Zhu, H., Chang, W., Zhou, Y., Li, Z., Chen, J., Pang, J., Shen, C., & He, T. (2026). π³: Permutation-Equivariant Visual Geometry Learning. International Conference on Learning Representations (ICLR).
We hope the above response can help address your concerns. We have also made targeted revisions to the manuscript based on your suggestions. We are happy to answer any additional questions you may have.
Dear Reviewer GepG,
Thank you for your thoughtful review and constructive suggestions, which have significantly improved the quality and clarity of our paper. We sincerely hope that our rebuttal has fully addressed all of your concerns. As the discussion period is ending soon, if you need any further clarification or additional information, please do not hesitate to contact us. We truly appreciate your time and effort devoted to reviewing our work.
Best regards, Authors of Submission 12062
Author Response - Part 1
Thanks for your thoughtful comments, questions and insightful suggestions. We provide detailed responses below.
W1 & Q2 - Novelty of our method
We thank the reviewer for the comment. The main novelty of IVGT lies in its formulation of implicit visual geometry learning, rather than in individual components. Our central idea is to infer a shared and continuously queryable 3D field from unposed images, extending visual geometry transformers beyond discrete and view-indexed predictions. This representation provides an amodal description of the scene and a common geometric basis for surface extraction, depth and normal rendering, and novel-view synthesis. The contribution therefore concerns what the model learns to represent and how its outputs are geometrically coupled. Feature sampling and SDF-based volume rendering provide established tools for realizing this idea. We position our contribution at the level of the scene representation and learning formulation, rather than claiming these tools as new.
Specifically, VGGT represents geometry with per-pixel points. IVGT changes both what the model represents and how its geometric outputs are decoded. IVGT provides continuous representation of the 3D scene including the occluded spaces, enabling tasks including surface extraction and novel-view rendering. We acknowledge that PixelNeRF and VolSDF establish important foundations for image-conditioned neural fields and SDF-based reconstruction. Multi-view PixelNeRF assumes known relative camera poses, while VolSDF requires per-scene optimization. However, it is not trivial to implement our formulation (generalizable, continous, and consistent scene reconstruction from unposed images) directly with existing techniques. Challenges include noisy estimated poses, consistent geometry requirement, good generalization ability, simultanously apperance and geometry modeling, etc. In order to address them, we have designed and tried various ways of feature aggregation, positional encoding, training strategies, training objectives, etc., to build a model that achieves strong performance on comprehensive tasks, including mesh and point cloud reconstruction, novel view synthesis, depth and surface normal estimation, and camera pose estimation.
W2 & Q1 - Comparison with fast per-scene optimization baselines
Thanks for the suggestion. Among the per-scene optimization baselines in Table 1 (in our paper), MonoSDF is the only one for which IVGT performs better on some metrics and worse on others. Given that the original MonoSDF requires approximately 22 hours of per-scene optimization, we believe IVGT still offers a substantial advantage. Nevertheless, we agree with the reviewer that comparisons with faster per-scene optimization baselines would make the evaluation more complete. As shown in Table (a), we additionally compare IVGT with (1) the Instant-NGP variant described in the MonoSDF paper, (2) Neuralangelo[1], (3) 2DGS[2], and (4) QGS[3]. We also report the average optimization time measured when reproducing each method.
Table (a). Comparison with fast per-scene optimization baselines.
IVGT achieves substantially better reconstruction quality while requiring much less time than these faster per-scene optimization methods.
[1]Li Z, Müller T, Evans A, et al. Neuralangelo: High-fidelity neural surface reconstruction[C]//2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2023: 8456-8465.
[2]Huang B, Yu Z, Chen A, et al. 2d gaussian splatting for geometrically accurate radiance fields[C]//ACM SIGGRAPH 2024 conference papers. 2024: 1-11.
[3]Zhang Z, Huang B, Jiang H, et al. Quadratic gaussian splatting: High quality surface reconstruction with second-order geometric primitives[C]//2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025: 28260-28270.