VGGT-R: You Can Really Accelerate
Vision Geometry Grounded Transformer
with the Underestimated Registers

Xuefeng Jiang*, Zibin Geng*, Yuan Ma*, Jia Li, Peijin Jia, Sheng Sun,
Wenke Huang, Guoping Pan, Yi Cheng, Yuheng Wang, Liujiang Yan†, Ke Wang†

KargoBot · University of Chinese Academy of Sciences · Zerith Robotics
Tsinghua University · Nanyang Technological University

* Equal contribution † Corresponding author

▤ Paper ⌘ Code “ BibTeX

Interactive 360-degree reconstruction of the 7Scenes Chess scene

Drag to rotate

Interactive 360-degree reconstruction of the 7Scenes Red Kitchen scene

Drag to rotate

Abstract

Corresponding input image sequence and VGGT-R geometric results

VGGT-R repurposes pretrained register tokens as compact cross-frame proxies, accelerating long-sequence 3D perception without retraining or explicit token pruning.

Large-scale 3D vision foundation models such as Visual Geometry Grounded Transformer (VGGT) deliver strong performance across geometric perception tasks, but the quadratic complexity of global attention severely limits long-sequence scalability. We present VGGT-R, a training-free framework that repurposes register tokens to reduce global attention overhead while preserving geometric performance. A fine-grained layer-wise analysis reveals that registers act as compact cross-frame proxies. We split the network into a geometric front-end with restricted attention and a geometric back-end with register-guided token selection. VGGT-R achieves 3.2× speedup without explicitly pruning tokens, while improving depth estimation by approximately 10% and remaining competitive across other geometric tasks.

Method

A layer-aware design that gives registers a different job at each stage of geometric reasoning.

Restricted Attention

Cross-frame communication is routed solely through camera and register tokens in selected front-end blocks, eliminating dense patch-to-patch interactions.

Register-Guided Selection

Three complementary signals preserve relevant, structurally distinctive and cross-frame novel patches in the back-end keys and values.

Training-Free Integration

The original token layout and prediction heads remain unchanged. No additional data, fine-tuning or explicit patch-token pruning is required.

Results

VGGT-R improves the accuracy–efficiency trade-off across depth, point clouds and camera pose estimation.

3.2×speedup
at 1,000 frames

0.2143depth RMSE
at 1,000 frames

0.976normal consistency
on NRGBD

0extra training
or token pruning

Depth Estimation on ScanNet-50

Frames	Method	AbsRel ↓	RMSE ↓	RMSE-log ↓	δ1 ↑	δ3 ↑
100	VGGT	0.0707	0.2576	0.1353	0.9289	0.9917
	FastVGGT	0.0691	0.2504	0.1331	0.9308	0.9923
	VGGT-R (Ours)	0.0664	0.2380	0.1264	0.9370	0.9929
500	VGGT	0.0698	0.2525	0.1334	0.9294	0.9923
	FastVGGT	0.0665	0.2535	0.1293	0.9346	0.9932
	VGGT-R (Ours)	0.0662	0.2333	0.1243	0.9384	0.9934

Best results are highlighted. VGGT-R remains stable from 20 to 1,000 input frames and further improves as more views become available.

Visualization

Qualitative comparison of dense depth, reconstruction quality and register-guided token scoring.

Depth Estimation

VGGT-R reduces depth errors around object boundaries and challenging regions while producing more coherent geometry.

3D Point Cloud Reconstruction

Dense reconstructions retain fine-grained spatial information without explicit patch-token pruning.

Register-Guided Token Selection

The three cues capture complementary semantic, structural and temporal information.

BibTeX

@misc{jiang2027vggtr,
  title  = {VGGT-R: You Can Really Accelerate Vision Geometry Grounded
            Transformer with the Underestimated Registers},
  author = {Jiang, Xuefeng and Geng, Zibin and Ma, Yuan and Li, Jia
            and Jia, Peijin and Sun, Sheng and Huang, Wenke and others},
  year   = {2027}
}

VGGT-R: You Can Really AccelerateVision Geometry Grounded Transformerwith the Underestimated Registers