# TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention

Source: [TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention](<https://www.sota2.com/research/paper/6a06a8a6dbc836b87d82a15a>)

Collection index: [llms.txt](<https://www.sota2.com/research/llms.txt>)

Data updated: 2026-05-16T05:00:05.655Z

Authors: David Huang, Guile Wu, Chengjie Huang, Bingbing Liu, Dongfeng Bai

Published: 2026-05-14T03:24:09.000Z

## Abstract

Recent feed-forward 3D reconstruction methods, such as visual geometry transformers, have substantially advanced the traditional per-scene optimization paradigm by enabling effective multi-view reconstruction in a single forward pass. However, most existing methods struggle to achieve a balance between reconstruction quality and computational efficiency, which limits their scalability and efficiency. Although some efficient visual geometry transformers have recently emerged, they typically use the same sparsity ratio across layers and frames and lack mechanisms to adaptively learn representative tokens to capture global relationships, leading to suboptimal performance. In this work, we propose TurboVGGT, a novel approach that employs an efficient visual geometry transformer with adaptive alternating attention for fast multi-view 3D reconstruction. Specifically, TurboVGGT employs an end-to-end trainable framework with adaptive sparse global attention guided by adaptive sparsity selection to capture global relationships across frames and frame attention to aggregate local details within each frame. In the adaptive sparse global attention, TurboVGGT adaptively learns representative tokens with varying sparsity levels for global geometry modeling, considering that token importance varies across frames, attention layers operate tokens at different levels of abstraction, and global dependencies rely on structurally informative regions. Extensive experiments on multiple 3D reconstruction benchmarks demonstrate that TurboVGGT achieves fast multi-view reconstruction while maintaining competitive reconstruction quality compared with state-of-the-art methods. Project page: https://turbovggt.github.io/.

## Recorded benchmark results

A research result applies to its stated dataset, split, metric and evaluation configuration. Compare like-for-like settings, preserve the metric's unit and direction, and distinguish paper-reported results from independent reproduction. The best result recorded here is not a claim of exhaustive, current global coverage.

- [Point cloud reconstruction on 7-Scenes](<https://www.sota2.com/research/sota/point-cloud-reconstruction-on-7-scenes.md>): Point Cloud Reconstruction; 7 Scenes; Inference Time (s): 2.

- [Point-cloud reconstruction on ScanNet](<https://www.sota2.com/research/sota/point-cloud-reconstruction-on-scannet.md>): Point Cloud Reconstruction; ScanNet; Chamfer Distance: 40.

- [Depth Estimation on Sintel (single)](<https://www.sota2.com/research/sota/depth-estimation-on-sintel-single.md>): Depth Estimation; Sintel (single); AbsRel: 28.7.

- [Multi-view 3D Reconstruction on 7-Scenes dense setting](<https://www.sota2.com/research/sota/multi-view-3d-reconstruction-on-7-scenes-dense-setting.md>): Multi-view 3D Reconstruction; 7-Scenes dense setting; Peak Inference Memory (GB): 23.47.

- [Depth Estimation on 7-Scenes (dense)](<https://www.sota2.com/research/sota/depth-estimation-on-7-scenes-dense.md>): Depth Estimation; 7-Scenes dense; AbsRel: 0.296.

- [Camera Pose Estimation on N-RGBD](<https://www.sota2.com/research/sota/camera-pose-estimation-on-n-rgbd.md>): Camera pose estimation; N-RGBD; RRA@30: 100.

- [Depth Estimation on N-RGBD dense](<https://www.sota2.com/research/sota/depth-estimation-on-n-rgbd-dense.md>): Depth Estimation; N-RGBD dense; AbsRel: 1.3.

- [Camera Pose Estimation on 7-Scenes (RRA/RTA/AUC Metrics)](<https://www.sota2.com/research/sota/camera-pose-estimation-on-7-scenes-rra-rta-auc-metrics.md>): Camera pose estimation; 7 Scenes; RRA@30: 100.

- [Point Cloud Reconstruction on N-RGBD (Acc, Comp, NC Mean/Median, Time)](<https://www.sota2.com/research/sota/point-cloud-reconstruction-on-n-rgbd-acc-comp-nc-mean-median-time.md>): Point Cloud Reconstruction; N-RGBD; Accuracy (Mean): 2.5.

- [Video Depth Estimation on Sintel](<https://www.sota2.com/research/sota/video-depth-estimation-on-sintel.md>): Video Depth Estimation; Sintel; Delta Threshold Accuracy (1.25): 71.6.

- [Camera Pose Estimation on RealEstate10K](<https://www.sota2.com/research/sota/camera-pose-estimation-on-realestate10k.md>): Camera pose estimation; RealEstate10K; AUC@30: 84.31.

- [Video Depth Estimation on Bonn (AbsRel, δ \< 1.25)](<https://www.sota2.com/research/sota/video-depth-estimation-on-bonn-absrel-1-25.md>): Video Depth Estimation; BONN; AbsRel: 5.3.

- [3D Reconstruction on NRGBD (Mean/Median Acc, Comp, NC)](<https://www.sota2.com/research/sota/3d-reconstruction-on-nrgbd-mean-median-acc-comp-nc.md>): 3D Reconstruction; NRGBD; Accuracy Mean: 2.1.

## Original sources and implementation

[Original paper](<https://arxiv.org/abs/2605.14315>)

[PDF](<https://arxiv.org/pdf/2605.14315.pdf>)
