- 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
-
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.
-
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.
-
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
-
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.
-
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.
-
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.
-
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.
-
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
- 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.
- 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
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!
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 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 (512×512 pixel area) and HD inputs (1024×1024 pixel area), with varying aspect ratios. Importantly, the 2K FHD (1920×1080) and 2.5K QHD (2560×1440) 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 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.
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.