acceptodds
Under review as a conference paper at ICLR 2027

ReDT: Resolution-Decoupled Visual Geometry Transformer

Abstract

We introduce ReDT, a feed-forward visual geometry transformer designed to scale multi-view reconstruction to high-definition (HD) inputs. Existing visual geometry transformers are typically trained at fixed, standard definition (SD) resolutions and exhibit degraded reconstruction quality and poor computational scaling to high input resolutions. While recent efficient architectures reduce the complexity of inter-frame attention, expensive within-frame and backbone attention remain largely unaddressed, making HD reconstruction impractical with existing architectures. ReDT addresses both limitations with a resolution-aware architecture designed to scale to high-resolution inputs. The key idea of ReDT is to first map each view to a fixed-size set of basis tokens, compute correspondences in this compact constant-size basis space, and then project the outputs back to the original resolution for dense prediction. By decoupling the internal representation from input resolution during attention, ReDT substantially reduces the computational cost of high-resolution 3D reconstruction and improves model performance at all resolutions. Extensive experiments on multiple datasets spanning SD to FHD (2K) inputs show that ReDT outperforms state-of-the-art visual geometry transformers in both reconstruction quality and computational efficiency. For instance, on the ETH3D (2K) dataset, ReDT achieves a speedup, a % reduction in peak GPU memory usage, and a % reduction in chamfer distance over VGGT-.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual Geometry Transformer, High-Definition Reconstruction, Multi-View Reconstruction

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_4kn2OpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper present ReDT, which aims to address the resolution sensitivity and the computational expense of feed-forward multi-view reconstruction models. ReDT first combines adaptive patch merging with area-weighted projection, then it utilizes test-time training to perform correspondence reasoning in the compact space of a fixed number of latent basis tokens. The method is instantiated on VGGT and VGGT-Ω and predicts multiple geometric quantities through the compact spatial feature via the prediction heads. The authors provide comprehensive evaluation that includes multi-view reconstruction benchmarks and the resolution sensitivity analyses. The component ablations are informative and easy to follow. The reported results show clear improvements in high-resolution reconstruction accuracy and efficiency, while retaining competitive performance at standard resolutions compared with previous methods.

Critical strengths
  1. ReDT achieves substantial improvements in both reconstruction accuracy and efficiency of high-resolution inputs. The core contribution of the paper is the support of high-resolution inputs while keeping the memory consumption affordable, where the paper demonstrates a practically improvement rather than the acceleration alone.

  2. The evaluation and ablation of the paper are comprehensive and easy to follow. The evaluation provides extensive results across several tasks and backbone variants. The results demonstrate strong HD results and competitive performance. In addition, the improvements with the original VGGT backbone also suggest that the benefits are not solely attributable to the stronger VGGT-Ω initialization.

  3. The paper has detailed explanation of the the architectural choices, which are also supported by analysis and component ablations. The authors identifies resolution-related degradation even when upsampling adds no visual information and also shows that removing basis projection worsens reconstruction and causes an out-of-memory failure at the highest tested resolution. These provide evidence to support the compact representation’s utility.

Critical weaknesses
  1. From the efficiency aspect, the scaling experiments do not fully establish the cost of the complete end-to-end reconstruction pipeline. In Figure2, the authors present the reconstruction performance excluding the task heads to support the claim that the bottleneck lies in the attention mechanism. Though this is highly informative, its nearly flat runtime cannot directly characterize end-to-end behavior. I suggest that the authors provide the corresponding full-pipeline measurements and a precise description of the timing scope either in the main paper or in the appendix, which would strengthen the practical scalability claim. In addition, I suggest mentioning the hardware for computing in Table. 1 for better tables presentation.

  2. The extent of generalization to unseen resolutions remains unclear. The authors provide mixed fine-tuning results in Appendix A.1 but do not specify the actual training dimensions. This makes it difficult to determine which evaluation resolutions represent extrapolation beyond training and how much robustness comes from architectural changes versus resolution coverage during fine-tuning. Since training a high-resolution model with relatively good performace for HD inputs should include high-resolution training data, I would suggest the authors further provide details in the corresponding section for a better demonstration.

  3. The ablation is detailed and clear, one suggestion would be that the contribution of area-aware weighting is not isolated from that of the latent bottleneck. Table 4 shows the results of removing the entire basis projection, while Appendix A.5 relies on consistency assumptions about features and positional representations. A matched uniform-weight ablation would better establish whether the proposed weighting mechanism specifically explains the observed robustness.

  4. For the experiments section, I suggest adding more qualitative comparison or qualitative results to demonstrate the high-resolution predictions, which can better show the high-resolution details of the prediction. This would make the motivation of the paper more clear and demonstrate the contribution more directly by visualization.

  5. There are some inconsistent details in the paper: a) Sparse/dense results are reversed in Section 4.2. The description of 7-Scenes does not match Table 3: ReDT obtains CD 0.033 for sparse sampling and 0.032 for dense sampling, while pi3 achieves the best sparse result of 0.032. b) The resolution labels in Appendix A.6 are inconsistent. Appendix A.6 describes 3840×2160 as “4K” in the dataset discussion but labels it “8K” in the resolution table; the same table labels 2560×1440 as “4K”.

Questions for authors
  1. The authors utilized TTT in the pipeline for high-resolution inputs and this is easy to follow itself, from my perspective, TTT is somehow related to linear attention, and several works in other fields like video generation uses linear attention for HD inputs. Utilizing linear attention or mixed attention like MLA in the Kimi model may have achieve the goal of HD inputs as well. I think this can be a future work or a discussion in the paper. This is just a suggestion and I don't need the authors to provide further experiments. I would appreciate the authors share their thoughts in the rebuttals.
  2. There are some previous methods that try to address high-resolution inputs by utilizing the position-interpolated RoPE. I think it is suitable to discuss these lines of work though they do not target on multiview solely. I would include some of the references here:

@inproceedings{dens3r, title={Dens3R: A Foundation Model for 3D Geometry Prediction}, author={Xianze Fang and Jingnan Gao and Zhe Wang and Zhuo Chen and Xingyu Ren and Jiangjing Lyu and Qiaomu Ren and Zhonglei Yang and Xiaokang Yang and Yichao Yan and Chengfei Lyu}, booktitle={The Fourteenth International Conference on Learning Representations}, year={2026} }

@inproceedings{ngo2026dage, title={DAGE: Dual-Stream Architecture for Efficient and Fine-Grained Geometry Estimation}, author={Ngo, Tuan Duc and Huang, Jiahui and Oh, Seoung Wug and Blackburn-Matzen, Kevin and Kalogerakis, Evangelos and Gan, Chuang and Lee, Joon-Young}, booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)}, year={2026} }

3 comments

We thank the reviewer for the valuable comments. We have revised the paper with the main changes marked in blue. We address the reviewer's concerns point by point below. Note that we have updated the resolution labels throughout the revised manuscript: the previously labeled 4K QHD and 8K UHD are now referred to as 2.5K QHD and 4K UHD, respectively. The actual input resolutions and experimental results remain unchanged.

Q3-1: The scaling experiments do not fully establish the cost of the end-to-end pipeline: Full-pipeline measurements and a precise description?

Response: Thanks for this suggestion. We clarify that the scaling experiments in Fig.2 measure inference time excluding task heads, aiming to isolate the computational cost of the backbone. To resolve this concern, we additionally provide full end-to-end pipeline latency and peak memory measurements in Appendix New Fig.10. The findings and conclusions drawn from Fig.2 remain consistent with the inclusion of full-pipeline measurements. In Fig.10, the runtime of ReDT-Ω\Omega slightly increases with increasing input resolution because the computational cost of task heads becomes more significant, contributing to the overall end-to-end latency. In addition, in the revised manuscript, we have updated Table 1 to clarify that all compared methods are measured on the same single data-center GPU system with approximately 140 GB of high-bandwidth memory under identical environmental settings and timing protocol.

Q3-2: Training resolutions and generalization to unseen resolutions?

Response: Apologies for the lack of clarity in the original manuscript. We have clarified the training resolution settings in Appendix A.1. Specifically, our mixed-resolution training uses SD inputs (512×512512\times512 pixel area) and HD inputs (1024×10241024\times1024 pixel area), with varying aspect ratios. Importantly, the 2K FHD (1920×10801920\times1080) and 2.5K QHD (2560×14402560\times1440) evaluation resolutions exceed the maximum training resolution by approximately 2$\times$ and 3.5$\times$ in pixel area, respectively. Therefore, these evaluations represent unseen resolution extrapolation without resolution-specific fine-tuning, rather than resolutions covered during training. As shown in Table 4, ReDT-Ω\Omega maintains comparable CD from 1K HD (0.1137) to 2K FHD (0.1140), with moderate degradation at 2.5K QHD (0.1249), demonstrating its generalization beyond the training resolution range.

Q3-3: Contribution of area-aware weighting?

Response: Thanks for this valuable suggestion. To address this concern, we conduct an additional matched ablation by replacing area-aware weighting with uniform weighting while keeping all other components unchanged. This result evaluates the effectiveness of area-aware weighting in improving reconstruction accuracy with high-resolution inputs. We have added this ablation and the corresponding discussion to the revised appendix.

ETH3D(HD)CD ↓AbsRel ↓AUC@30 ↑
ReDT-Ω\Omega (Uniform Weighting)0.1210.04284.9
ReDT-Ω\Omega (Area-aware Weighting)0.1140.03686.1

Q3-4: Additional qualitative comparisons?

Response: Thanks for this suggestion. We have added additional high-resolution qualitative comparisons in the revised manuscript. Specifically, Appendix New Figs.7-9 present cross-resolution reconstruction results on identical scenes and input views, along with close-up comparisons of fine geometric structures and object boundaries. These results further demonstrate ReDT's ability to leverage high-resolution information for recovering fine geometric details, complementing the quantitative results discussed in Reviewer nE42 Q1-1.

Q3-5: Some inconsistent details?: Sparse/dense results are reversed in Section 4.2? Inconsistent resolution labels?

Response: Thanks for the reminder.

  1. Sparse/dense results. We have corrected the description in Section 4.2 to match Table 3. Specifically, ReDT achieves CD values of 0.033 and 0.032 under sparse and dense settings, respectively, while π3\pi^3 achieves the best sparse result of 0.032.
  2. Resolution labels. We have corrected the inconsistent resolution labels in Appendix A.6. Specifically, we now consistently refer to 2560×14402560\times1440 as 2.5K QHD and 3840×21603840\times2160 as 4K UHD, and have updated the corresponding descriptions and tables accordingly. The actual input resolutions and experimental results remain unchanged.

Q3-6: Discussion of linear attention or mixed attention (MLA) for HD reconstruction?

Response: Thanks for this insightful suggestion. We agree that TTT is closely related to linear attention as certain TTT formulations can be interpreted as learned linear attention operators. Exploring linear attention or hybrid attention architectures such as Kimi Linear, which combines Kimi Delta Attention (KDA) with Multi-Head Latent Attention (MLA), is a promising direction for efficient HD reconstruction. However, directly replacing Scaled Dot-Product Attention (SDPA) with these efficient attention mechanisms may not necessarily preserve HD reconstruction quality, as high-resolution 3D reconstruction requires precise modeling of fine-grained interactions that may be approximated by these mechanisms. Further investigation is needed to balance computational efficiency and geometric accuracy. We have added a Future Work section discussing this direction in the revised manuscript.

Q3-7: Discussion of related methods that address high-resolution inputs by utilizing the position-interpolated RoPE?

Response: Thanks for bringing these related works to our attention. We have cited them in the revised Related Work section. For this line of work, such as Dens3R and DAGE, the focus is on improving the handling of high-resolution inputs through position-interpolated RoPE and related techniques. Specifically, Dens3R employs position-interpolated RoPE to improve robustness to high-resolution inputs, while DAGE combines interpolated RoPE with a dual-stream architecture to preserve fine geometric details. In contrast, ReDT focuses on resolution-decoupled geometric reasoning through DIVA, which projects variable-length spatial tokens into a fixed-size basis space to control computational complexity as input resolution increases. These approaches address complementary aspects of high-resolution reconstruction.

Thank you for the response and I appreciate the newly added experiments, especially the clarification and experinents for training resolutions and generalization to unseen resolutions and for the area-aware weighting. I also appreciate the discussion of future work and the authors' thoughts on linear attention or mixed attention. Personally, I think this is also a explorable direction to address the same resolution problem and I appreciate the explanation of why "directly replacing Scaled Dot-Product Attention (SDPA)" may not necessarily preserve the quality. Still, I believe this is a practical approach to address the resolution sensitivity and the computational expense of feed-forward multi-view reconstruction models, so I would keep my score. Good luck!

Reviewer_L4yQOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ReDT, a resolution decoupled visual geometry transformer. The paper starts from a deep analysis about the performance of recent SoTA methods on high resolution reconstruction. I personally like this part which forms a good starting point and motivation of this paper. The authors propose two method contributions: (1) Variable Patch Size Encoder and (2) Discretization Invariant Attention Blocks. The first one adaptively decreases the number of input tokens and the second one enables attention module processing with more tokens. Those are reasonable. Experimental results show a good performance compared with previous methods.

Critical strengths
  • The analysis in Section 3.2 is thoughtful and fairly comprehensive. It goes beyond merely runtime and memory measurements and provide useful insights into resolution sensitivity and feature redundancy. They give the proposed approach a clear motivation.
  • Building a more efficient visual geometry model for high-resolution inputs is a worthwhile direction, and the paper targets a relevant practical problem.
Critical weaknesses
  • The method section is difficult to follow. Several steps are introduced without enough explanation, and I found it hard to reconstruct the forward pass of a DIVA block from the current description. In particular, the TTT operation and the mapping back to spatial tokens need clarification. When a component is adopted from prior work, a brief explanation of how it works and how it is used here would help considerably.

  • The technical novelty of DIVA is not clearly established. Soft token assignment, normalized aggregation into a compact latent space, and latent-to-spatial mapping are closely related to existing approaches, while the TTT mechanism is adopted from ZipMap. The paper should explain more clearly what is new beyond combining these components and applying them to more parts of the network.

  • It is unclear how much of the improvement comes from high-resolution training. Were the proposed model and the baselines trained or fine-tuned under comparable resolution settings? A useful control would be to fine-tune baselines such as ZipMap and FastVGGT on the same data with the same resolution schedule. Without this comparison, it is difficult to separate architectural gains from gains due to training.

  • Some baseline results need further explanation. In particular, the large gap between VGGT and VGGT-Ω in camera pose AUC@30 in Table 2 is weird. This does not necessarily indicate an evaluation issue, but it would help to double check the experiments for each method. DA3 and Pi3 also somehow present unexpectedly bad results in the table.

  • The efficiency contribution of the adaptive tokenizer needs a more direct evaluation. The component-removal ablation is useful, but it does not show how many tokens are actually removed. Please report some quantative comprisions in this case and this would make its practical contribution much clearer.

Questions for authors

I would like to ask the authors to address my concerns in the Critical Weaknesses section. The main questions that could affect my assessment are:

  1. Could you clarify the forward pass of a DIVA block, including the TTT operation and the mapping back to spatial tokens?
  2. What are the specific technical contributions beyond the components adopted from prior work?
  3. Were the proposed model and the baselines trained under comparable data and resolution settings? How much of the improvement can be attributed to the architecture rather than high-resolution training?
4 comments

We thank the reviewer for the valuable comments. We have revised the paper with the main changes marked in blue. We address the reviewer's concerns point by point below. Note that we have updated the resolution labels throughout the revised manuscript: the previously labeled 4K QHD and 8K UHD are now referred to as 2.5K QHD and 4K UHD, respectively. The actual input resolutions and experimental results remain unchanged.

Q2-1: Method section explanation: Clarification of DIVA forward pass, TTT operation and the mapping back to spatial tokens?

Response: Thanks for the valuable question. In the revised manuscript, we have clarified the complete forward pass of DIVA in Section 3.3, including the adopted TTT mechanism and the mapping back to spatial tokens. Specifically, the DIVA-based reconstruction pipeline consists of four steps:

  1. Basis projection: DIVA computes token-to-basis assignment matrices AQA^Q and AKVA^{KV}, and projects query, key, and value features into KK fixed-size basis tokens using quadrature weights cic_i to account for the spatial area of each token.
  2. TTT-based reasoning: Following ZipMap, TTT performs feature interactions in the compact basis space through inner-loop updates of a temporary model state. Unlike ZipMap, ReDT applies TTT to the KK basis tokens rather than the full spatial token sequence.
  3. Spatial reconstruction: The resulting basis features are mapped back to the NN spatial tokens using the transpose of the spatial-to-basis projection operator, recovering spatial representations for subsequent processing.
  4. Task prediction: After the backbone, variable-size tokens are densified to spatial feature maps for the DPT decoder, while camera tokens are processed by the camera head.

Q2-2: Technical novelty of DIVA beyond existing components?

Response: Thanks for the suggestion. The novelty of DIVA lies in its resolution-decoupled formulation for multi-view geometric reconstruction, rather than these individual operations. Specifically, DIVA introduces two key designs:

  1. Quadrature-aware basis projection. Unlike conventional token aggregation, DIVA explicitly accounts for the spatial area of variable-size patches through quadrature weights, enabling a more consistent representation across different input discretizations, as discussed in Appendix A.5.
  2. Resolution-decoupled geometric reconstruction in a discretization-invariant space. Motivated by our effective-rank analysis, DIVA performs feature interactions in a fixed-size basis space throughout the image encoder and attention backbone, while mapping the resulting features back to spatial tokens for dense geometric prediction. Unlike ZipMap, which primarily accelerates global attention, this design also addresses the resolution-dependent computational costs of encoder and frame attention. The effectiveness of these designs is supported by Table 4, where removing basis projection substantially degrades reconstruction accuracy and causes out-of-memory failures at 2.5K (QHD) resolution. We have further clarified these technical distinctions and the contributions adopted from prior work in the revised manuscript.

Q2-3: Fair comparison under matched high-resolution training?

Response: Thanks for the reminder. We clarify that all benchmarked methods are fine-tuned using the same training protocol as ReDT, as described in Appendix A.1 and A.6. The training datasets only contain SD inputs (512×512512\times512) with 10–20 views and HD inputs (1024×10241024\times1024) with 2–4 views because training with higher-resolution inputs and more views would be computationally prohibitive. The 2K FHD (1920×10801920\times1080) and 2.5K QHD (2560×14402560\times1440) evaluation resolutions exceed the maximum training resolution and represent unseen resolution extrapolation without resolution-specific fine-tuning, rather than resolutions covered during training. Therefore, the improvements in Tables 1 and 2 are not attributed simply to additional high-resolution training data.

Q2-4: Some baseline results need further explanation: In Table 2, large gap between some methods?

Response: Thanks for raising this concern. We clarify the following points.

  1. Performance gap between VGGT and VGGT-Ω\Omega. Although all baselines have been fine-tuned, VGGT-Ω\Omega is a separately pretrained model with substantial architectural improvements and more extensive pretraining data, rather than a simple variant of VGGT. Thus, the performance gap between VGGT and VGGT-Ω\Omega reflects the potential impact of their different architectures and pretraining. Moreover, early stopping leads to fewer optimization steps for VGGT, which may further contribute to the gap.
  2. High-resolution generalization. All baselines retain their originally recommended hyperparameters without HD-specific tuning. Since the 2K/2.5K/4K evaluation resolution exceeds the training resolution range, differences in resolution sensitivity may further contribute to the observed performance gaps.
  3. DA3 and π3\pi^3 performance. DA3 actually achieves competitive AUC@30 scores of 75.28 and 79.22 on MVS-Synth and ScanNet++, respectively, outperforming VGGT-Ω\Omega on ScanNet++. For π3\pi^3, its relatively low AUC@30 on MVS-Synth (20.83) is associated with lower translation accuracy (RTA: 44.58), despite achieving 100.00 rotation accuracy (RRA). We have clarified these points in the revised manuscript.

Q2-5: Quantitative evaluation of adaptive tokenization efficiency: The efficiency contribution of the adaptive tokenizer needs a more direct evaluation?

Response: Thanks for this valuable suggestion. To address this concern, we additionally report the number of spatial tokens before and after adaptive tokenization across different input resolutions, together with the corresponding token reduction ratios. Combined with the accuracy ablation in Table 4, these results provide a more comprehensive evaluation of the adaptive tokenizer's contribution to reconstruction efficiency and quality. Notably, the token reduction ratio increases from 34% at 1K HD to 48% at 4K UHD, demonstrating greater spatial token savings at higher resolutions.

ETH3DOriginal TokensAdaptive TokensReduction (%)
1K HD3600237434%
2K FHD8040489339%
2.5K QHD14400816443%
4K UHD324001689448%

Dear Reviewer,

Thank you again for your constructive comments. We hope our rebuttal has addressed your concerns. Should any concerns remain, we would be grateful for the opportunity to provide further clarification or additional experimental evidence during the discussion period.

Thank you for your time and consideration.

Thanks for the rebuttal. Most of my concerns are addressed and I changed my rating to weak acceptance.

Reviewer_nE42OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents ReDT, a novel framework for efficient 3D reconstruction from high-resolution (HD) images. It combines several techniques, including adaptive tokenization, discretization-invariant attention, and test-time training, to adapt visual geometry transformers to process HD images, which would otherwise incur substantial computational costs with conventional transformer architectures. The method is evaluated on point-cloud reconstruction, camera pose estimation, and depth estimation across multiple datasets, demonstrating strong quantitative performance and computational efficiency.

Critical strengths
  1. Strong quantitative results and extensive evaluation. The quantitative results demonstrate superior performance compared to existing baselines on the high-resolution benchmarks, supported by extensive experiments across multiple datasets and geometric reconstruction tasks. These results provide evidence of the effectiveness and computational efficiency of the proposed framework.

  2. Well-motivated architecture and efficiency analysis. Figures 2 and 3 demonstrate favorable computational scaling and relatively stable reconstruction performance as input resolution increases. The analysis of resolution sensitivity and effective rank provides useful motivation for the proposed architecture. Additionally, the theoretical complexity analysis in the appendix further supports the efficiency of the proposed method.

Critical weaknesses
  1. Insufficient justification for high-resolution reconstruction and limited qualitative evaluation. The paper does not convincingly demonstrate why HD inputs are preferable to conventional standard-definition (SD) inputs. For example, Table 4 shows almost no improvement in Chamfer Distance from 1K to 2K, while performance degrades at 4K. Therefore, the practical benefits of using HD images remain unclear. More analysis and experiments are needed to address this concern. Additionally, the qualitative results are limited to a few examples in the appendix, which do not fully demonstrate the benefits of HD reconstruction. It would be helpful to include cross-resolution visual comparisons, especially close-up evaluations of fine geometric details to demonstrate whether the method can effectively exploit the additional information in HD images. This is a major concern because high-resolution reconstruction is the central motivation of the paper.

  2. Incomplete comparisons with relevant baselines. Although the paper compares against standard geometry backbones such as VGGT and π³, as well as several efficient variants, it lacks comparisons with methods specifically designed for high-resolution reconstruction, such as HD-VGGT [1]. Furthermore, the paper does not establish whether ReDT-HD provides meaningful advantages over existing methods operating at their standard lower resolutions. For example, comparing ReDT-HD with VGGT-Ω-SD in terms of reconstruction quality, runtime, and memory would help demonstrate its practical benefits. Although the appendix specifies that all baselines are fine-tuned using a unified mixed-resolution training protocol, this does not fully address the question of whether evaluating existing methods at HD resolutions is representative of their standard operating configurations.

[1] T. Chen et al., “HD-VGGT: High-Resolution Visual Geometry Transformer,” arXiv preprint arXiv:2603.27222, 2026. https://arxiv.org/abs/2603.27222

Questions for authors
  1. What are the practical benefits of using HD images for reconstruction? HD images potentially contain more high-frequency information and fine geometric details than SD images. However, can ReDT effectively exploit this additional information to improve reconstruction quality? Could the authors provide quantitative evidence demonstrating these benefits?

  2. How are existing baselines adapted to high-resolution inputs, and is the comparison fair? Methods such as VGGT are typically designed to operate at much lower resolutions. Are the baselines simply evaluated with resized HD inputs, or are additional architectural modifications applied? How does ReDT-HD compare against existing methods operating at their standard resolutions in terms of accuracy, runtime, and memory?

  3. Can the authors provide more qualitative comparisons demonstrating the benefits of HD reconstruction? In particular, cross-resolution comparisons on the same scenes, with close-ups of fine geometric structures and reconstruction boundaries, would help demonstrate whether the proposed method benefits from the additional spatial information in HD inputs.

  4. Can the authors clarify the absence of comparisons with HD-VGGT? Given that HD-VGGT also targets efficient high-resolution reconstruction and is discussed in the related work, a direct comparison would help establish the advantages of ReDT over existing HD-specific methods.

3 comments

We thank the reviewer for the valuable comments. We have revised the paper with the main changes marked in blue. We address the reviewer's concerns point by point below. Note that we have updated the resolution labels throughout the revised manuscript: the previously labeled 4K QHD and 8K UHD are now referred to as 2.5K QHD and 4K UHD, respectively. The actual input resolutions and experimental results remain unchanged.

Q1-1: Insufficient justification for high-resolution reconstruction: What are the practical benefits of using HD images for reconstruction? Quantitative evidence?

Response: Thanks for this valuable comment.

  1. Clarification of Table 4. The 2K (FHD) and 2.5K (QHD) resolutions in Table 4 are unseen resolution extrapolations beyond the training range, and therefore are not necessarily expected to show monotonic improvements in CD. As described in Appendix A.1, our mixed-resolution training uses SD inputs (512×512512\times512 pixel area) and HD inputs (1024×10241024\times1024 pixel area) (apologies for the lack of clarity in the original manuscript). Table 4 demonstrates that ReDT-Ω\Omega maintains comparable CD at the unseen 2K (FHD) resolution and shows only slight degradation at the unseen 2.5K (QHD) resolution. These results highlight ReDT's ability to generalize beyond its training resolution range, rather than suggesting that increasing resolution necessarily improves global reconstruction accuracy. We have clarified this in the revised manuscript.
  2. Benefits of high-resolution reconstruction. High-resolution images preserve fine-grained spatial information that may be lost through downsampling, particularly for thin structures, object boundaries, and local surface details. To demonstrate these benefits, we additionally compare ReDT-Ω\Omega using SD and HD inputs on ETH3D. Here, high-detail CD evaluates reconstruction errors in ground-truth high-curvature regions. HD inputs reduce the overall CD by 4% and the high-detail CD by 12%, demonstrating their benefits for recovering fine geometric structures.
ETH3DOverall CD ↓High-detail CD ↓
ReDT-Ω\Omega SD0.1190.162
ReDT-Ω\Omega HD0.114 (↓4%)0.142 (↓12%)

Q1-2: Fair comparisons with standard baselines: How are existing baselines adapted to high-resolution inputs? How does ReDT-HD compare against baselines operating at standard resolutions?

Response: Thanks for this question.

  1. Baseline for high-resolution inputs. As described in Appendix A.1 and A.6, all baselines are fine-tuned from their publicly available pretrained checkpoints using the same training datasets and mixed-resolution training protocol as ReDT. For fair comparisons, all methods are evaluated on the same high-resolution input images under the same evaluation protocol and environment.
  2. Comparison with baselines operating at standard resolutions. Following the reviewer's suggestion, we additionally compare ReDT-Ω\Omega at HD resolutions against VGGT-Ω\Omega operating at its standard SD resolution, reporting reconstruction quality, runtime, and memory usage. It is worth noting that comparing the inference cost of ReDT at high resolutions against baselines at their standard resolutions may not be a directly comparable setting because the inference cost can significantly increase with input resolution and ReDT primarily focuses on enabling efficient high-resolution reconstruction rather than demonstrating that HD reconstruction must be better than SD reconstruction. Nevertheless, the existing results in Table 1 demonstrate that ReDT-Ω\Omega achieves lower CD, faster inference, and lower peak memory than VGGT-Ω\Omega at the same resolution.
ETH3DCD ↓High-detail CD ↓Time (s) ↓Memory (GB) ↓
VGGT-Ω\Omega SD0.1200.1640.217.40
ReDT-Ω\Omega HD0.114 (↓5%)0.142 (↓13%)1.0112.40
VGGT-Ω\Omega 2K0.1700.241102.7098.09
ReDT-Ω\Omega 2K0.114 (↓33%)0.148 (↓39%)6.3564.47

Q1-3: Limited qualitative comparisons: Can the authors provide more qualitative comparisons demonstrating the benefits of HD reconstruction?

Response: Thanks for this suggestion. We additionally provide cross-resolution qualitative comparisons in the revised manuscript. Specifically, Appendix New Figs.7-9 present reconstruction results of VGGT-Ω\Omega at SD resolution, ReDT-Ω\Omega at SD resolution, and ReDT-Ω\Omega at HD resolution, along with close-up comparisons of fine geometric structures and object boundaries. These comparisons highlight ReDT's ability to leverage high-resolution spatial information for recovering fine geometric details and preserving reconstruction boundaries, complementing the quantitative results in Q1-1.

Q1-4: Incomplete comparisons with relevant baselines: Can the authors clarify the absence of comparisons with HD-VGGT?

Response: Thanks for this question.

  1. We have attempted to reimplement HD-VGGT for comparison. However, HD-VGGT involves a complex multi-stage pipeline with specific configurations. Consequently, our reimplementation did not yield reliable results for a fair comparison.
  2. We have discussed the technical differences between HD-VGGT and ReDT in Lines 122–127 of the original manuscript. Specifically, HD-VGGT adopts a dual-branch architecture combining low-resolution global geometric reasoning with high-resolution feature refinement, whereas ReDT performs geometric reasoning in a discretization-invariant space through DIVA, without requiring a separate high-resolution refinement branch. We have further clarified the differences from HD-VGGT in the revised manuscript.
  3. ReDT primarily focuses on improving computational efficiency and resolution scalability instead of explicitly enhancing fine-grained reconstruction accuracy. We therefore compare against standard geometry backbones and efficient SOTA methods under the same evaluation protocol, as reported in Tables 1 and 2.

Thanks for the response. I think the newly added experiments are helpful to determine the uesfulness of using HD images in reconstruction and the contribution of the paper. Therefore, I will keep my original score. I suggest putting the qualitative results in the main text for completeness and better demonstration of the model.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.