pith. sign in

arxiv: 2606.18439 · v1 · pith:KAZASH4Jnew · submitted 2026-06-16 · 💻 cs.CV · cs.RO

RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer

Jinhao You (1) , Shuo Lyu (1) , Zhuohang Lyu (1) , Tanxuan Li (1) , Zibo Zhao (1) , Jiaxiang Hu (2) , Kai Tang (3) , Yichen Guo (3) ((1) University of Pennsylvania
show 3 more authors
(2) University of California Irvine (3) Nanyang Technological University)
This is my paper

Pith reviewed 2026-06-27 01:03 UTC · model grok-4.3

classification 💻 cs.CV cs.RO
keywords 3D reconstructiontransformerattention compressionmulti-view geometryredundancy removalvisual geometry grounded transformer
0
0 comments X

The pith

RegimeVGGT compresses VGGT layer by layer to achieve 6.7 times faster 3D scene reconstruction at the same quality.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows that VGGT's layers fall into three distinct regimes with different roles in 3D reconstruction. Shallow layers have little cross-view structure, middle layers handle alignment, and deep layers are redundant for geometry but needed for pose. By applying different compression strategies to each regime using saliency-guided merging and protected downsampling, it removes redundancy while keeping performance. This matters because it makes dense 3D recovery from images much faster without retraining, addressing the quadratic attention cost that limits scalability.

Core claim

RegimeVGGT identifies three regimes in VGGT via spectral, probing, and causal analyses: shallow layers lack cross-view structure, middle layers drive cross-view alignment, and deep layers are redundant for dense geometry yet their cross-frame attention is essential for pose. It then applies layer-wise U-shaped compression with Saliency-Guided Banded Merging to protect salient tokens and Selectively Protected K/V Downsampling to preserve spatial coverage and pose-critical paths, resulting in 6.7x speedup over VGGT* at matched reconstruction quality.

What carries the argument

Layer-wise U-shaped compression using Saliency-Guided Banded Merging and Selectively Protected K/V Downsampling, guided by the three identified regimes in attention layers.

If this is right

  • Training-free acceleration allows deployment on resource-limited devices for multi-view 3D tasks.
  • Preservation of pose estimation alongside geometry enables accurate camera registration in compressed models.
  • Uniform compression is avoided by tailoring to layer heterogeneity, improving efficiency over prior accelerators.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar regime analysis could apply to other multi-view or video transformers to find compressible layers.
  • The method might extend to other dense prediction tasks like optical flow if regimes are analogous.

Load-bearing premise

The three regimes in VGGT layers remain consistent across different scenes and datasets, and the saliency and phase-shifted protection methods retain all necessary information for geometry and pose without new errors.

What would settle it

Running RegimeVGGT on a held-out dataset and observing a significant drop in reconstruction metrics like accuracy or pose error compared to the original VGGT would falsify the claim of matched quality at speedup.

Figures

Figures reproduced from arXiv: 2606.18439 by (2) University of California, (3) Nanyang Technological University), Irvine, Jiaxiang Hu (2), Jinhao You (1), Kai Tang (3), Shuo Lyu (1), Tanxuan Li (1), Yichen Guo (3) ((1) University of Pennsylvania, Zhuohang Lyu (1), Zibo Zhao (1).

Figure 1
Figure 1. Figure 1: Overview of RegimeVGGT. VGGT’s 24 global-attention layers expose a three-band rank structure: shallow (L1–L10) and deep (L19–L24) bands have effective rank ≤ 20, while the middle band (L11–L18) peaks at ≈ 90 where cross-view alignment is concentrated. The same band partition configures two orthogonal training-free operators: Saliency-Guided Banded Merging compresses the token-count axis with a U-shaped per… view at source ↗
Figure 2
Figure 2. Figure 2: Heterogeneous geometry tokens and middle-band concentration. (a) Ablating the top-10% patch tokens by squared depth residual (after regressing out 2D cues) disrupts the depth head 2.1× more than random ablation and more than a Sobel-ranked image-edge baseline on 3D-rich scenes; the gap narrows on flat scenes (Flower), confirming the score captures 3D boundaries rather than 2D texture. Bottom panels overlay… view at source ↗
Figure 3
Figure 3. Figure 3: Predicted camera trajectories on ScanNet-50 scene0648_01 under 1000-frame infer￾ence. VGGT baseline (left), RegimeVGGT (center), and FastVGGT (right) compared against ground truth (gray); predictions colored by per-frame ATE. Large-scale outdoor pose [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Three-band layer structure in VGGT’s cross-frame attention. Top: for a fixed query patch in Frame 0, attention to all patch tokens in Frame 1 and Frame 2, visualized as spatial heatmaps. Bottom: cross-frame attention submatrix from all patch tokens in Frame 0 to all patch tokens in Frame 1. A.1.1 Layer-wise Probe of Geometric Information Closed-form ridge probe. At each global-attention layer ℓ, we ask how… view at source ↗
Figure 5
Figure 5. Figure 5: Geometric probe gain emerges in the middle band L11–L18. Signed normalized gain ∆(ℓ)/ maxj |∆j | across VGGT’s 24 aggregator layers for three dense probes (pointmap, depth, surface normal). Solid curves: default configuration with G2F removed. Dashed curves: boundary￾scan variants retaining G2F through L1–K. Shaded region: the critical band L11–L18. A.1.2 Causal and Structural Diagnostics of Cross-View Ali… view at source ↗
Figure 6
Figure 6. Figure 6: Four diagnostics co-localize cross-view alignment to L11–L18. Rows: 1−norm. entropy (sharpness), diagonal ratio (cross-frame correspondence), causal influence on depth head, causal influence on camera head; each row per-row normalized. The critical band L11–L18 (red box) is bright on all four rows; the camera CI peaks near L15. A.1.3 Rank Spectrum of Global Attention Target-free structural diagnostic. The … view at source ↗
Figure 7
Figure 7. Figure 7: Effective attention rank is universally inverted-U-shaped, peaking at the middle band L11–L18. Effective rank of the global-attention matrix vs. layer depth, on five scenes (TNT Courthouse, TNT Barn, TNT Truck, DTU, ScanNet); m=256 sampled query rows per layer. Shallow (ℓ≤10) and deep (ℓ≥19) layers concentrate in the top singular component (eff-rank/m<0.05); the middle band ℓ∈[11, 18] peaks above the 0.30 … view at source ↗
Figure 8
Figure 8. Figure 8: DINOv2 [CLS] attention is edge-aware but not edge-only, motivating its use as the protection signal in RegimeVGGT. Across five scenes (rows): (col 1) source frame with the top-1 CLS-attention patch outlined in cyan; (col 2) DINO CLS attention map from the last patch-embed block, with the top-5% patches (the geometry tokens we protect) outlined in cyan; (col 3) cosine similarity between the query patch and … view at source ↗
Figure 9
Figure 9. Figure 9: Additional visualizations of pose estimation results on the ScanNet dataset. [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
read the original abstract

Visual Geometry Grounded Transformer (VGGT) recovers dense 3D scene structure from multi-view images in one forward pass, but quadratic cross-frame attention limits its scalability. Existing training-free accelerators reduce computation uniformly along one axis, missing layer heterogeneity. Our spectral, probing, and causal analyses reveal three regimes: shallow layers lack cross-view structure, middle layers drive cross-view alignment, and deep layers are redundant for dense geometry yet their cross-frame attention remains essential for pose. RegimeVGGT applies layer-wise U-shaped compression along two axes: Saliency-Guided Banded Merging protects geometry- and edge-salient tokens, while Selectively Protected K/V Downsampling preserves cross-frame spatial coverage and the pose-critical path through a phase-shifted spatial grid, a reference-frame anchor, and uncompressed camera/register tokens. Training-free, RegimeVGGT achieves a 6.7x speedup over VGGT* at matched reconstruction quality.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces RegimeVGGT, a training-free acceleration for the Visual Geometry Grounded Transformer (VGGT) used in one-pass dense 3D scene reconstruction from multi-view images. Spectral, probing, and causal analyses identify three layer regimes (shallow: no cross-view structure; middle: cross-view alignment; deep: redundant for geometry but pose-critical). Layer-wise U-shaped compression is applied along two axes—Saliency-Guided Banded Merging (protecting geometry/edge-salient tokens) and Selectively Protected K/V Downsampling (phase-shifted grid, reference-frame anchor, uncompressed camera/register tokens)—yielding a claimed 6.7x speedup over VGGT* at matched reconstruction quality.

Significance. If the regime boundaries prove stable and the protection rules preserve all tokens required for both dense geometry and pose, the work would offer a practical, training-free route to scaling quadratic-attention geometry transformers by exploiting layer heterogeneity rather than uniform pruning. The dual-axis, spatially preserving design and explicit reference-frame anchoring are concrete strengths that could be adopted in related efficient-ViT efforts for 3D vision.

major comments (2)
  1. [Abstract] Abstract: the central 6.7x speedup at matched reconstruction quality is asserted without any supporting quantitative numbers, error bars, dataset statistics, ablation tables, or controls, which is load-bearing for the empirical claim.
  2. [Abstract] Abstract (regime analyses): the three regimes are presented as stable enough for fixed layer-wise rules to never drop essential tokens, yet no cross-dataset regime-consistency metrics or ablation removing the reference-frame anchor are supplied; if regime transitions shift with scene scale or lighting, the matched-quality guarantee is at risk.
minor comments (2)
  1. [Abstract] The baseline VGGT* is referenced without an explicit definition or citation; clarify whether it denotes the original model, a re-implementation, or a modified variant.
  2. [Abstract] The abstract would benefit from naming the concrete metrics (e.g., accuracy, completeness, pose error) used to declare 'matched reconstruction quality'.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for noting the potential practical value of exploiting layer heterogeneity in geometry transformers. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the central 6.7x speedup at matched reconstruction quality is asserted without any supporting quantitative numbers, error bars, dataset statistics, ablation tables, or controls, which is load-bearing for the empirical claim.

    Authors: We agree that the abstract would benefit from concrete quantitative anchors for the primary claim. In the revised manuscript we will insert a concise statement of the 6.7x speedup together with the corresponding reconstruction-quality metric (mean error) and the number of scenes/views used for the main result. The full supporting evidence—error bars, per-dataset statistics, ablation tables, and controls—already appears in Sections 4 and 5; the abstract change simply makes this evidence visible at the summary level. revision: yes

  2. Referee: [Abstract] Abstract (regime analyses): the three regimes are presented as stable enough for fixed layer-wise rules to never drop essential tokens, yet no cross-dataset regime-consistency metrics or ablation removing the reference-frame anchor are supplied; if regime transitions shift with scene scale or lighting, the matched-quality guarantee is at risk.

    Authors: The regime boundaries were derived from spectral, probing, and causal analyses performed on the primary multi-view evaluation sets; the protection mechanisms (saliency-guided banded merging, phase-shifted grid downsampling, reference-frame anchoring, and uncompressed camera/register tokens) were explicitly introduced to keep geometry- and pose-critical tokens intact even if boundaries shift modestly. We nevertheless recognize that explicit cross-dataset regime-consistency statistics and a dedicated ablation that removes the reference-frame anchor would strengthen the stability argument. Both will be added in the revision. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical outcome of compression rules

full rationale

The paper presents the 6.7x speedup as the measured result of applying fixed layer-wise rules (Saliency-Guided Banded Merging and phase-shifted K/V downsampling) after identifying regimes via spectral/probing/causal analyses. No equations, fitted parameters, or self-citations are shown that would make the speedup equivalent to its inputs by construction. The central claim remains an independent empirical outcome rather than a renaming or self-referential definition.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Abstract-only review; no free parameters, ad-hoc axioms, or invented entities are stated. The method rests on standard transformer attention and multi-view geometry assumptions already present in the VGGT baseline.

axioms (2)
  • standard math Standard multi-head attention and positional encoding mechanics in vision transformers
    Invoked implicitly as the substrate for the compression rules.
  • domain assumption Multi-view geometry provides consistent cross-frame correspondences that can be protected by saliency and spatial-grid rules
    Underlies the claim that protected tokens suffice for pose and geometry.

pith-pipeline@v0.9.1-grok · 5761 in / 1388 out tokens · 50558 ms · 2026-06-27T01:03:54.163245+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

42 extracted references · 14 canonical work pages · 6 internal anchors

  1. [1]

    Building rome in a day.Communications of the ACM, 54(10):105–112, 2011

    Sameer Agarwal, Yasutaka Furukawa, Noah Snavely, Ian Simon, Brian Curless, Steven M Seitz, and Richard Szeliski. Building rome in a day.Communications of the ACM, 54(10):105–112, 2011

    2011

  2. [2]

    Neural rgb-d surface reconstruction

    Dejan Azinovi´c, Ricardo Martin-Brualla, Dan B Goldman, Matthias Nießner, and Justus Thies. Neural rgb-d surface reconstruction. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6290–6301, 2022

    2022

  3. [3]

    Token merging for fast stable diffusion

    Daniel Bolya and Judy Hoffman. Token merging for fast stable diffusion. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4599–4603, 2023

    2023

  4. [4]

    Token merging: Your ViT but faster

    Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. InInternational Conference on Learning Representations (ICLR), 2023

    2023

  5. [5]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020

    2020

  6. [6]

    Scannet: Richly-annotated 3d reconstructions of indoor scenes

    Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017

    2017

  7. [7]

    FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning

    Tri Dao. Flashattention-2: Faster attention with better parallelism and work partitioning.arXiv preprint arXiv:2307.08691, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  8. [8]

    Vision Transformers Need Registers

    Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers.arXiv preprint arXiv:2309.16588, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  9. [9]

    Vision meets robotics: The kitti dataset.The international journal of robotics research, 32(11):1231–1237, 2013

    Andreas Geiger, Philip Lenz, Christoph Stiller, and Raquel Urtasun. Vision meets robotics: The kitti dataset.The international journal of robotics research, 32(11):1231–1237, 2013

    2013

  10. [10]

    Which tokens to use? investigating token reduction in vision transformers

    Joakim Bruslund Haurum, Sergio Escalera, Graham W Taylor, and Thomas B Moeslund. Which tokens to use? investigating token reduction in vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 773–783, 2023

    2023

  11. [11]

    Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera

    Shahram Izadi, David Kim, Otmar Hilliges, David Molyneaux, Richard Newcombe, Pushmeet Kohli, Jamie Shotton, Steve Hodges, Dustin Freeman, Andrew Davison, et al. Kinectfusion: real-time 3d reconstruction and interaction using a moving depth camera. InProceedings of the 24th annual ACM symposium on User interface software and technology, pages 559–568, 2011

    2011

  12. [12]

    Large scale multi-view stereopsis evaluation

    Rasmus Jensen, Anders Dahl, George V ogiatzis, Engin Tola, and Henrik Aanæs. Large scale multi-view stereopsis evaluation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 406–413, 2014

    2014

  13. [13]

    Token fusion: Bridging the gap between token pruning and token merging

    Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. InProceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1383–1392, 2024

    2024

  14. [14]

    Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

    Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: Bench- marking large-scale scene reconstruction.ACM Transactions on Graphics (ToG), 36(4):1–13, 2017

    2017

  15. [15]

    Video token merging for long-form video understanding.arXiv preprint arXiv:2410.23782, 2024

    Seon-Ho Lee, Jue Wang, Zhikang Zhang, David Fan, and Xinyu Li. Video token merging for long-form video understanding.arXiv preprint arXiv:2410.23782, 2024

    work page arXiv 2024

  16. [16]

    Grounding image matching in 3d with mast3r

    Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with mast3r. InEuropean conference on computer vision, pages 71–91. Springer, 2024

    2024

  17. [17]

    S-vggt: Structure-aware subscene decomposition for scalable 3d foundation models.arXiv preprint arXiv:2603.17625, 2026

    Xinze Li, Pengxu Chen, Yiyuan Wang, Weifeng Su, and Wentao Cheng. S-vggt: Structure-aware subscene decomposition for scalable 3d foundation models.arXiv preprint arXiv:2603.17625, 2026. 10

    work page arXiv 2026

  18. [18]

    Not all patches are what you need: Expediting vision transformers via token reorganizations.arXiv preprint arXiv:2202.07800, 2022

    Youwei Liang, Chongjian Ge, Zhan Tong, Yibing Song, Jue Wang, and Pengtao Xie. Not all patches are what you need: Expediting vision transformers via token reorganizations.arXiv preprint arXiv:2202.07800, 2022

    work page arXiv 2022

  19. [19]

    Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras.IEEE transactions on robotics, 33(5):1255–1262, 2017

    Raul Mur-Artal and Juan D Tardós. Orb-slam2: An open-source slam system for monocular, stereo, and rgb-d cameras.IEEE transactions on robotics, 33(5):1255–1262, 2017

    2017

  20. [20]

    DINOv2: Learning Robust Visual Features without Supervision

    Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V o, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision.arXiv preprint arXiv:2304.07193, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  21. [21]

    Vision transformers for dense prediction

    René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12179–12188, October 2021

    2021

  22. [22]

    Dynamicvit: Efficient vision transformers with dynamic token sparsification.Advances in neural information processing systems, 34:13937–13949, 2021

    Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification.Advances in neural information processing systems, 34:13937–13949, 2021

    2021

  23. [23]

    Structure-from-motion revisited

    Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion revisited. InProceed- ings of the IEEE conference on computer vision and pattern recognition, pages 4104–4113, 2016

    2016

  24. [24]

    Pixelwise view selection for unstructured multi-view stereo

    Johannes L Schönberger, Enliang Zheng, Jan-Michael Frahm, and Marc Pollefeys. Pixelwise view selection for unstructured multi-view stereo. InEuropean conference on computer vision, pages 501–518. Springer, 2016

    2016

  25. [25]

    FastVGGT: Training-Free Acceleration of Visual Geometry Transformer

    You Shen, Zhipeng Zhang, Yansong Qu, Xiawu Zheng, Jiayi Ji, Shengchuan Zhang, and Liujuan Cao. Fastvggt: Training-free acceleration of visual geometry transformer.arXiv preprint arXiv:2509.02560, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    Scene coordinate regression forests for camera relocalization in rgb-d images

    Jamie Shotton, Ben Glocker, Christopher Zach, Shahram Izadi, Antonio Criminisi, and Andrew Fitzgibbon. Scene coordinate regression forests for camera relocalization in rgb-d images. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 2930–2937, 2013

    2013

  27. [27]

    Litevggt: Boosting vanilla vggt via geometry-aware cached token merging

    Zhijian Shu, Cheng Lin, Tao Xie, Wei Yin, Ben Li, Zhiyuan Pu, Weize Li, Yao Yao, Xun Cao, Xiaoyang Guo, et al. Litevggt: Boosting vanilla vggt via geometry-aware cached token merging. arXiv preprint arXiv:2512.04939, 2025

    work page arXiv 2025

  28. [28]

    Avggt: Rethinking global attention for accelerating vggt.arXiv preprint arXiv:2512.02541, 2025

    Xianbing Sun, Zhikai Zhu, Zhengyu Lou, Bo Yang, Jinyang Tang, Liqing Zhang, He Wang, and Jianfu Zhang. Avggt: Rethinking global attention for accelerating vggt.arXiv preprint arXiv:2512.02541, 2025

    work page arXiv 2025

  29. [29]

    Probabilistic robotics.Communications of the ACM, 45(3):52–57, 2002

    Sebastian Thrun. Probabilistic robotics.Communications of the ACM, 45(3):52–57, 2002

    2002

  30. [30]

    Bold features to detect texture- less objects

    Federico Tombari, Alessandro Franchi, and Luigi Di Stefano. Bold features to detect texture- less objects. InProceedings of the IEEE international conference on computer vision, pages 1265–1272, 2013

    2013

  31. [31]

    Block-Sparse Global Attention for Efficient Multi-View Geometry Transformers

    Chung-Shien Brian Wang, Christian Schmidt, Jens Piekenbrinck, and Bastian Leibe. Faster vggt with block-sparse global attention.arXiv preprint arXiv:2509.07120, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Vggt: Visual geometry grounded transformer

    Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 5294–5306, 2025

    2025

  33. [33]

    Continuous 3d perception model with persistent state

    Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A Efros, and Angjoo Kanazawa. Continuous 3d perception model with persistent state. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 10510–10522, 2025

    2025

  34. [34]

    Dust3r: Geometric 3d vision made easy

    Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 20697–20709, 2024. 11

    2024

  35. [35]

    Httm: Head-wise temporal token merging for faster vggt.arXiv preprint arXiv:2511.21317, 2025

    Weitian Wang, Lukas Meiner, Rai Shubham, Cecilia De La Parra, and Akash Kumar. Httm: Head-wise temporal token merging for faster vggt.arXiv preprint arXiv:2511.21317, 2025

    work page arXiv 2025

  36. [36]

    π3: Scalable permutation-equivariant visual geometry learning.arXiv e-prints, pages arXiv–2507, 2025

    Yifan Wang, Jianjun Zhou, Haoyi Zhu, Wenzheng Chang, Yang Zhou, Zizun Li, Junyi Chen, Jiangmiao Pang, Chunhua Shen, and Tong He. π3: Scalable permutation-equivariant visual geometry learning.arXiv e-prints, pages arXiv–2507, 2025

    2025

  37. [37]

    Multi- modal token fusion for vision transformers

    Yikai Wang, Xinghao Chen, Lele Cao, Wenbing Huang, Fuchun Sun, and Yunhe Wang. Multi- modal token fusion for vision transformers. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12186–12195, 2022

    2022

  38. [38]

    Dymu: Dynamic merging and virtual unmerging for efficient vlms.arXiv preprint arXiv:2504.17040, 2025

    Zhenhailong Wang, Senthil Purushwalkam, Caiming Xiong, Silvio Savarese, Heng Ji, and Ran Xu. Dymu: Dynamic merging and virtual unmerging for efficient vlms.arXiv preprint arXiv:2504.17040, 2025

    work page arXiv 2025

  39. [39]

    Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass

    Jianing Yang, Alexander Sax, Kevin J Liang, Mikael Henaff, Hao Tang, Ang Cao, Joyce Chai, Franziska Meier, and Matt Feiszli. Fast3r: Towards 3d reconstruction of 1000+ images in one forward pass. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 21924–21935, 2025

    2025

  40. [40]

    MVSNet: Depth inference for unstructured multi-view stereo

    Yao Yao, Zixin Luo, Shiwei Li, Tian Fang, and Long Quan. MVSNet: Depth inference for unstructured multi-view stereo. InEuropean Conference on Computer Vision (ECCV), pages 767–783, 2018

    2018

  41. [41]

    Spargeattention: Accurate and training-free sparse attention accelerating any model inference.arXiv preprint arXiv:2502.18137, 2025

    Jintao Zhang, Chendong Xiang, Haofeng Huang, Jia Wei, Haocheng Xi, Jun Zhu, and Jianfei Chen. Spargeattention: Accurate and training-free sparse attention accelerating any model inference.arXiv preprint arXiv:2502.18137, 2025

    work page arXiv 2025

  42. [42]

    Streaming 4D Visual Geometry Transformer

    Dong Zhuo, Wenzhao Zheng, Jiahe Guo, Yuqi Wu, Jie Zhou, and Jiwen Lu. Streaming 4d visual geometry transformer.arXiv preprint arXiv:2507.11539, 2025. 12 A Extended Analysis Evidence This appendix collects the empirical evidence underlying RegimeVGGT’s two design axes—token merging and K/V downsampling—together with the experiments that motivated each desi...

    work page internal anchor Pith review Pith/arXiv arXiv 2025