- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
This paper introduces TraDE, a training-free token reduction method designed to accelerate global attention in visual geometry transformers, including VGGT, Pi-3 and VGGT-Omega. TraDE formulates token reduction as an iterative clustering process, merging clusters within local spatio-temporal windows when the resulting increase in cosine-based distortion remains below a predefined threshold. This enables adaptive token reduction based on input characteristics rather than a fixed token budget. During global attention, only one representative token per cluster is processed, and its output is subsequently propagated to the remaining cluster members. Cluster assignments are updated at manually selected stage boundaries. Experiments demonstrate speedups of up to 7.25×, 8.05×, and 4.47× for VGGT, Pi-3 and VGGT-Omega respectively, on 1000-frame ScanNet sequences, with additional evaluations on 7Scenes, NRGBD, ETH3D and TUM-Dynamics.
Critical strengths
- The developed method is training-free and is shown on three backbones so it is plug-and-play.
- The target task is important and useful to a lot of down-streamed applications.
- The authors evaluate widely across different backbone networks and different datasets.
Critical weaknesses
- Table 4's gains come mainly from repartitioning, which keeps ≈39% of tokens in the middle layers vs. ≈14% elsewhere. A fixed per-stage ratio with the same schedule is the missing control. Without it, the gains cannot be credited to input-adaptivity.
- TraDE is evaluated at a single operating point against baselines at their defaults. Moreover, "Ret." is a linear average while attention cost is quadratic. Specifically, at the "same" 26.69% retention, TraDE runs at 15.46 FPS vs. 18.4 FPS for the fixed variants. Accuracy-vs-latency Pareto curves are needed.
- On ScanNet, TraDE has the lowest Pi-3 AUC@3 of all methods at 300-1000 frames (e.g., 17.89 vs. 23.57 dense), which is the setting behind the "8.0×" headline. On NRGBD, AUC@3 falls from 83.60 to 67.80, showing a single \lambda does not transfer across backbones.
- The conceptual novelty appears somewhat limited. The objective resembles DP-means/facility location, while the merging criterion reduces to a size-weighted cosine threshold. Stage-wise repartitioning and spatio-temporal search are also related to AVGGT and HTTM. Several relevant works are missing, including DP-means, ToMe, ToMeSD, Agglomerative Token Clustering and Run-Length Tokenization. Although input-adaptive token merging may be new to geometric foundation models, the broader token-merging literature is already extensive, making the methodological novelty less compelling.
- Hand-crafted design choices weaken the central premise. Stage boundaries are manually selected for each backbone and noticeably affect performance (AUC@3: 14.42 vs. 15.28). Hyperparameters also appear to be tuned on 7Scenes, which is used for evaluation. Moreover, repartitioning at every layer performs substantially worse, suggesting that cluster distortion may not be a reliable proxy for geometric error.
- I think the authors miss some relevant baselines. AVGGT (CVPR 2026) supplies the stage analysis TraDE builds on and is training-free on VGGT/Pi-3, reporting ≈8-10× at 800 frames on 7Scenes, yet it is not compared. LiteVGGT (CVPR 2026) uses the same ScanNet-50/7Scenes/NRGBD protocol with released code, and at least its training-free cached-merging variant should be compared. The same applies to HeSS (CVPR 2026), the head-adaptive successor of the SparseVGGT baseline already used. Spark3R (arXiv 2605.06270, evaluated on VGGT-Ω) and GoToHunt (arXiv 2605.23892) should at least be discussed and hopefully be compared against.
- Reproduced baselines deviate from published numbers. Under FastVGGT's own ScanNet-50 protocol (Table 8), FastVGGT's published ≈10% CD gain over VGGT* at 1000 frames (0.471→0.425; LiteVGGT reports 0.485→0.436) vanishes (0.467→0.469). ATE and ARE are also systematically off, while RPE matches the FastVGGT paper almost exactly. FastVGGT's measured speedups fall below every published figure (7Scenes 1.87× vs. 2.74–3.65×; ScanNet-50 at 1000 frames 3.39× vs. 4.0–4.9×), and these gaps all favor TraDE.
Questions for authors
Please see the weakness
9 comments
Response to Weaknesses
W4. We thank the reviewer for highlighting these connections. We provide a detailed clarification of our methodological contributions below and will incorporate the key distinctions into the revised manuscript. We acknowledge that our cluster-merging objective follows the penalized-clustering principle of DP-means and facility-location formulations, while agglomerative clustering has also been explored in ATC. However, our contribution extends beyond these existing principles to an efficient constrained agglomerative merging algorithm for visual geometry Transformers. Specifically, TraDE evaluates the aggregate distortion change of participating clusters to guide representative selection, parallel merging of disjoint pairs, and adaptive termination, with incremental cluster statistics enabling efficient updates. More importantly, these mechanisms are integrated into a unified training-free framework that adaptively determines both which tokens to retain and how many, according to the evolving token representations across inputs and network depth.
Regarding AVGGT and HTTM, we have explicitly acknowledged their influence in the manuscript (Line 209 and Lines 215-216). Although AVGGT reveals the stage-wise characteristics of geometric Transformers, it still relies on predefined structured K/V subsampling, whereas TraDE determines token retention from representation redundancy through distortion-guided clustering. HTTM performs merging within predefined spatio-temporal blocks, while TraDE dynamically inherits candidate relations during merging, allowing clusters to progressively extend beyond their initial neighborhoods (Lines 215–216). TraDE shares the fundamental principle underlying ToMe, ToMeSD, and Run-Length Tokenization: exploiting redundancy among similar visual tokens to reduce computation. However, TraDE differs from these methods in its specific token-reduction procedure and the target models and tasks it addresses. While Run-Length Tokenization exploits pixel-level temporal redundancy before Transformer inference, TraDE performs adaptive clustering on intermediate token representations. Our contribution is not merely input-adaptive token merging, but a distortion-guided clustering procedure that accommodates heterogeneous redundancy across input sequences, spatial regions, and network depth. This enables clusters with different sizes and representatives to emerge within the same input, while stage-wise repartitioning further updates both cluster assignments and retention levels as representations evolve.
W5. Thank you for these comments. We clarify the rationale behind our stage-wise design, hyperparameter selection, and cluster distortion. First, our stage boundaries follow AVGGT's stage-wise observations, further supported by Appendix Fig. 9, which shows distinct attention patterns across network depth. Uniform repartitioning may reuse token partitions across stages despite changes in token representations, whereas our stage-wise strategy updates partitions at these transitions, explaining its superior performance in Table 7. Second, Appendix Table 5 and Fig. 5 demonstrate favorable accuracy–efficiency trade-offs across different values. Increasing from 0.04 to 0.05/0.06 improves acceleration from 3.95× to 4.44×/4.88×, while maintaining competitive AUC@30 (79.03 vs. 78.88/78.70). We select as a balanced configuration and consistently apply the same hyperparameters across all datasets and backbones. Finally, cluster distortion measures representation discrepancy rather than downstream geometric error, as explicitly acknowledged in Appendix B.3 (Lines 861–871). Therefore, repartitioning at every layer does not necessarily improve geometric accuracy. Our stage-wise strategy instead balances adaptation to evolving representations against computational overhead.
Response to Weaknesses
W6. Thank you for highlighting these relevant methods. Our original baseline selection considered both methodological relevance and implementation availability. At the time of the original experiments, we had not obtained official implementations of AVGGT, HeSS, or Spark3R. Moreover, the complete LiteVGGT pipeline incorporates fine-tuning and FP8 quantization, differing from our training-free setting. Given the extensive recent literature and limited space, we initially selected representative baselines covering token merging and sparse attention. Nevertheless, we agree that broader comparisons would strengthen our evaluation.
To address this concern, we have reproduced AVGGT and conducted additional comparisons with AVGGT, LiteVGGT, and GoToHunt, as reported in the Appendix Table 14. Specifically, AVGGT is reproduced with , while LiteVGGT is evaluated without fine-tuning or FP8 quantization to ensure a fair training-free comparison. On 7Scenes, TraDE achieves the highest inference efficiency (3.95×, 15.46 FPS), compared with 3.30×, 3.41×, and 3.50× for AVGGT, LiteVGGT, and GoToHunt, respectively. TraDE also achieves the highest AUC@3 (15.28) among these accelerated methods while maintaining competitive reconstruction quality. Notably, GoToHunt's reported runtime excludes MegaLoc preprocessing.
We will also expand the Related Work section to discuss all the suggested methods. Since AVGGT, LiteVGGT, and Spark3R have already been discussed in the main text, we briefly clarify the remaining methods below. HeSS allocates head-wise sparse-attention budgets based on offline sensitivity calibration, whereas TraDE performs distortion-guided token clustering to determine retention directly from current representations without additional offline calibration. GoToHunt combines diversity-based reference-frame selection with layer-dependent K/V subsampling, relying on predefined compression budgets. In contrast, TraDE adaptively determines representative tokens and retention ratios through cluster-distortion-guided merging, while stage-wise repartitioning accommodates evolving representations across network depth, without requiring an external place-recognition model or offline profiling.
W7. Thank you for raising this concern. We re-examined the FastVGGT evaluation in Appendix Table 8. All three methods were evaluated on identical scene inputs using the same metric definitions and timing procedure. Our measurements on an NVIDIA RTX PRO 6000 GPU show that FastVGGT and DenseVGGT took 84.86 s and 287.37 s per scene on average, respectively, yielding a 3.39× speedup. Differences in GPU hardware and runtime environments likely contribute to the discrepancy with the originally reported speedup. For geometric evaluation, our CD computation is consistent with FastVGGT's officially released evaluation code, following the same procedure for full-cloud bounding-box alignment, point sampling, voxel downsampling, and clipped bidirectional distances, without any modifications to these evaluation steps. For ATE and ARE, our evaluation logs confirm that Appendix Table 8 consistently uses Sim(3) alignment for all three methods. We further verified that EVO 1.37.0 rejects the simultaneous similarity and origin alignment requested by FastVGGT's released script, which may explain the discrepancies in ATE and ARE relative to its published results. Importantly, these discrepancies are likely attributable to differences in evaluation environments and dependencies, rather than inconsistent evaluation among the compared methods. All methods in our experiments use the same official VGGT checkpoint, identical inputs, and a unified evaluation protocol. Therefore, despite differences from the absolute results reported in FastVGGT, our comparisons remain fair and controlled, providing a reliable assessment of TraDE's accuracy–efficiency trade-off. Furthermore, we have already provided our implementation and evaluation code in the supplementary material, enabling independent verification and reproduction of our reported results.
Response to Weaknesses
W1. We thank the reviewer for the careful examination of our ablation studies. First, we emphasize that the token-similarity heterogeneity identified in Sec. 3.1 includes not only input-wise non-uniformity but also depth-wise dynamics. Accordingly, the higher retention ratio in intermediate layers (≈39% versus ≈14% in other stages) is itself a manifestation of TraDE's adaptivity, rather than a manually specified allocation. While stage-wise repartitioning determines when token partitions are updated, distortion-guided adaptive stopping determines how many tokens to retain according to the current representations. Moreover, Table 4 demonstrates the benefit of cluster-level distortion, improving AUC@3 from 13.88 to 14.28 over cosine-based pairing at the same 26.69% retention. To further isolate the contribution of adaptive retention, we introduce ΔE-Stage-Fixed in Appendix Table 13, which employs the same distortion-based clustering and stage-wise repartitioning as TraDE, but uses fixed retention ratios matched to TraDE's average retention at each stage. Under the same overall retention ratio of 26.69%, TraDE improves AUC@3/15/30 from 15.04/63.41/78.81 to 15.28/63.81/79.03, with comparable throughput (15.46 vs. 15.85 FPS). These results demonstrate that input-adaptive retention provides additional accuracy benefits beyond stage-wise repartitioning and predefined depth-dependent token allocation.
W2. We thank the reviewer for the valuable suggestion regarding computational efficiency and the fairness of performance comparisons. First, we acknowledge that the average token retention ratio (Ret.) does not directly reflect computational cost, since attention complexity depends quadratically on the number of retained tokens, while repartitioning also introduces additional overhead. Nevertheless, we report Ret. as an intuitive measure of token reduction. Second, TraDE has already been evaluated at multiple operating points in Appendix Table 5, where varying the trade-off coefficient λ demonstrates its flexibility in balancing geometric accuracy and inference efficiency. Furthermore, we have added an accuracy–efficiency trade-off curve in Appendix Fig. 5, comparing TraDE across different λ values with existing baselines on 7Scenes. The results demonstrate that TraDE achieves a more favorable accuracy–efficiency trade-off across a wide range of operating points. For instance, at λ=0.04, TraDE achieves higher AUC@3 than FastVGGT (15.28 vs. 14.98) while providing substantially greater acceleration (3.95× vs. 1.87×). Similarly, at λ=0.10, TraDE achieves comparable AUC@3 to HTTM (12.97 vs. 12.92) with significantly higher acceleration (5.59× vs. 2.92×). These comparisons further substantiate the effectiveness of TraDE beyond a single operating point.
W3. We thank the reviewer for carefully examining the results across different backbones and datasets. As an inference acceleration method, TraDE should be evaluated in terms of both geometric accuracy and computational efficiency. We acknowledge that, without any backbone-specific hyperparameter adjustments, TraDE achieves lower AUC@3 than FastVGGT on π³. However, this accuracy trade-off is accompanied by substantial efficiency gains. On ScanNet with 300–1000 input frames, TraDE exhibits an average relative AUC@3 reduction of 8.48% compared with FastVGGT, while achieving 2.04× higher inference speed and running 1.48× faster than HTTM, the fastest competing method. At 1000 frames, these advantages increase to 2.53× and 1.62×, respectively. Moreover, Table 2 shows that TraDE maintains competitive point-cloud reconstruction quality, with differences from FastVGGT of at most 0.001 in Acc/Comp and 0.003 in NC across these settings.
Similarly, on NRGBD, TraDE reduces AUC@3 from 83.60 to 67.80 relative to dense π³ (18.90%), while achieving 5.96× acceleration, corresponding to 2.11× and 1.56× higher inference speed than FastVGGT and HTTM, respectively. Meanwhile, AUC@30 decreases by only 2.25% relative to dense π³, and the point-cloud reconstruction metrics remain competitive. These results demonstrate a favorable accuracy–efficiency trade-off, although the degradation in AUC@3 also indicates that a unified λ may favor more aggressive acceleration on certain backbone–dataset combinations. To further investigate this trade-off, we conducted an additional experiment on π³ using a smaller λ=0.02 on NRGBD, as reported in Appendix Table 12. Reducing λ from 0.04 to 0.02 improves AUC@3 from 67.80 to 77.55 and AUC@30 from 95.66 to 97.00, while maintaining a 3.96× inference speedup.