ReDT: Resolution-Decoupled Visual Geometry Transformer

Abteen Arab; Guile Wu; Dongfeng Bai

cs.CVOct 18, 2026

ReDT: Resolution-Decoupled Visual Geometry Transformer

Authors: Abteen Arab, Guile Wu, Dongfeng Bai

Organizations: Huawei Noah’s Ark Lab · University of British Columbia

Abstract

We introduce ReDT, a feed-forward visual geometry transformer designed to scale multi-view reconstruction to high-definition (HD) inputs. Existing visual geometry transformers are typically trained at fixed, standard definition (SD) resolutions and exhibit degraded reconstruction quality and poor computational scaling to high input resolutions. While recent efficient architectures reduce the complexity of inter-frame attention, expensive intra-frame and encoder attention remain largely unaddressed, making HD reconstruction impractical with existing architectures. ReDT addresses both limitations with a resolution-aware architecture designed to scale to high-resolution inputs. The key idea of ReDT is to first map each view to a fixed-size set of basis tokens, compute correspondences in this compact constant-size basis space, and then project the outputs back to the original resolution for dense prediction. By decoupling the internal representation from input resolution during attention, ReDT substantially reduces the computational cost of high-resolution 3D reconstruction and improves model performance at all resolutions. Extensive experiments on multiple datasets spanning SD to FHD (2K) show that ReDT outperforms state-of-the-art visual geometry transformers in both reconstruction quality and computational efficiency. For instance, on the ETH3D (2K) dataset, ReDT achieves a $16\times$ speedup, a $34\%$ reduction in peak GPU memory usage, and a $33\%$ reduction in chamfer distance over VGGT- $Ω$ . Project Website.

Figures & tables

Figure 1: ReDT achieves state-of-the-art point-cloud reconstruction on ETH3D using 100 input views at 2K resolution (left), while enabling fast inference across HD benchmarks (right).

Figure 2: Reconstruction performance on the 20-view ETH3D (4K) dataset across input resolutions. (a-b) Runtime and peak GPU memory without task heads. (c) Reconstruction quality (CD). Pixel dimensions for each resolution are provided in the appendix.

Figure 3: All assessments are on the ETH3D dataset. (a) Reconstruction error (CD) under upsampling without added visual information. (b) Mean 1D Wasserstein distance from base-resolution features across VGGT- $\Omega$ frame-attention layers, with curves for different resolutions. (c) Effective rank (99% energy) of DINO, frame-attention, and global-attention features across resolutions.

Figure 4: ReDT architecture with adaptive tokenization, DIVA blocks, and dense prediction heads.

Method	SAIL-VOS 3D (1K)			ScanNet++ (1K)			UnrealStudios4K (2K)			ETH3D (2K)
Method	CD $\downarrow$	Time (s)	Mem. (GB)	CD $\downarrow$	Time (s)	Mem. (GB)	CD $\downarrow$	Time (s)	Mem. (GB)	CD $\downarrow$	Time (s)	Mem. (GB)
VGGT (CVPR 2025)	23.0525	59.00	47.25	1.0481	58.90	43.25	9.8623	116.88	101.72	2.5660	118.02	102.24
VGGT- $\Omega$ (CVPR 2026)	11.8681	28.67	36.77	0.1362	28.67	38.42	1.0299	102.66	93.24	0.1702	102.70	98.09
$\pi^{3}$ (ICLR 2026)	12.7053	55.52	37.94	0.3649	55.45	33.56	3.5208	109.46	98.57	0.7768	108.79	101.54
DepthAnything3 (ICLR 2026)	14.3069	20.90	46.77	0.1081	20.89	39.44	2.5909	80.91	97.82	0.3158	80.88	102.80
FastVGGT (ICLR 2026)	23.0125	21.32	34.11	0.8946	21.31	28.77	8.3933	84.29	94.88	2.4710	84.30	98.51

Table 1: Point-cloud reconstruction comparison across four high-resolution datasets. Lower Chamfer Distance (CD), runtime, and peak memory are better. Best results are bold , second-best results are underlined , and third-best results are double-underlined .

Method	Camera Pose Estimation						Depth Estimation
	MVS-Synth (2K)			ScanNet++ (2K)			DynamicReplica (1K)		Spring (2K)
	AUC@30 $\uparrow$	RTA $\uparrow$	RRA $\uparrow$	AUC@30 $\uparrow$	RTA $\uparrow$	RRA $\uparrow$	AbsRel $\downarrow$	$\boldsymbol{\delta<1.25}\ \boldsymbol{\uparrow}$	AbsRel $\downarrow$	$\boldsymbol{\delta<1.25}\ \boldsymbol{\uparrow}$
VGGT	14.16	40.11	88.21	5.04	36.80	44.74	0.1748	0.7812	0.3856	0.4473
VGGT- $\Omega$	83.46	98.80	100.00	69.99	93.33	100.00	0.0246	0.9938	0.0692	0.9457
$\pi^{3}$	20.83	44.58	100.00	44.98	79.70	95.67	0.0655	0.9702	0.2586	0.5586
DepthAnything3	75.28	94.30	100.00	79.22	98.28	99.65	0.0260	0.9944	0.1638	0.7267

Table 2: Camera pose estimation on MVS-Synth (2K) and ScanNet++ (2K), and depth estimation on DynamicReplica (1K) and Spring (2K). All methods use the fixed-HD evaluation protocol.

Method	ScanNet-100 Frames		ScanNet-300 Frames		ScanNet-500 Frames		7-Scenes (Sparse)		7-Scenes (Dense)
Method	CD $\downarrow$	Time (s)	CD $\downarrow$	Time (s)	CD $\downarrow$	Time (s)	CD $\downarrow$	Time (s)	CD $\downarrow$	Time (s)
VGGT	0.421	4.9	0.429	33.6	0.442	90.1	0.118	4.5	0.115	38.1
VGGT- $\Omega$	0.500	2.0	0.531	13.0	0.548	33.7	0.034	1.9	0.033	14.6
$\pi^{3}$	0.471	3.6	0.480	25.5	0.479	67.4	0.032	1.6	0.033	28.7
DepthAnything3	0.461	1.6	0.516	9.9	0.524	25.1	0.065	1.5	0.064	11.0
FastVGGT	0.416	2.8	0.430	12.3	0.437	28.4	0.109	2.7	0.115	14.2

Table 3: Point-cloud reconstruction comparison across standard-definition datasets. Lower CD and time are better. For time, best values are bold, second-best values are underlined, and third-best values are double-underlined.

Method	ETH3D (1K)			ETH3D (2K)			ETH3D (4K)
Method	CD $\downarrow$	AbsRel $\downarrow$	AUC@30 $\uparrow$	CD $\downarrow$	AbsRel $\downarrow$	AUC@30 $\uparrow$	CD $\downarrow$	AbsRel $\downarrow$	AUC@30 $\uparrow$
ReDT- $\Omega$ (Full)	0.1137	0.0356	86.09	0.1140	0.0332	86.82	0.1249	0.0425	85.75
$-$ Adaptive Tokenizer	0.1239	0.0644	84.16	0.1241	0.0627	84.74	0.1455	0.0943	83.77
$-$ Basis Projection	0.1344	0.0859	82.99	0.1631	0.1236	78.87	OOM	OOM	OOM
$-$ Test-Time Training	0.1195	0.0373	81.93	0.1198	0.0348	82.30	0.1313	0.0447	81.50

Table 4: Component-ablation for ReDT on ETH3D at three resolutions. Lower CD and AbsRel are better; higher AUC@30 is better. ’–’ refers to removing the component.

$K_{\mathrm{enc}}$	$K_{\mathrm{agg}}$	CD $\downarrow$	AbsRel $\downarrow$	AUC@30 $\uparrow$
512	64	0.1137	0.0356	86.09
256	64	0.3737	0.0367	36.18
512	128	0.1420	0.0430	81.79

Table 5: Ablation of encoder and aggregator basis counts.

Figure 5: Qualitative comparison on a dynamic Spider-Man sequence, showing depth maps, estimated camera trajectories, and reconstructed point clouds.

Figure 6: Qualitative comparison on a static ETH3D scene, showing depth maps, estimated camera trajectories, and reconstructed point clouds.

3D ViTs

3D Vision Transformers model spatial and volumetric data for tasks such as reconstruction, segmentation, detection, and pose estimation.

See all 145 papers

HD-VGGT: High-Resolution Visual Geometry Transformer
Mar 28, 2026Tianrun Chen, Yuanqi Hu, Yidong Han +113D Reconstruction 3D ViTs
KOKONI 3D, Moxin Technology · Zhejiang University · Huzhou University +4
TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention
May 14, 2026David Huang, Guile Wu, Chengjie Huang +23D ViTs Multi-View 3D Reconstruction
Huawei Noah’s Ark Lab · University of Toronto · Foundation Model Department, Huawei
VGGT-Prime: Compute-Adaptive Mixture-of-Heads for Efficient Visual Geometry Transformers
Sep 20, 2026Abteen Arab, Guile Wu, Chengjie Huang +13D ViTs Multi-View 3D Reconstruction
Huawei Noah’s Ark Lab · University of British Columbia

Mar 28, 2026J/K move · Enter open · S save

HD-VGGT: High-Resolution Visual Geometry Transformer

Tianrun Chen, Yuanqi Hu, Yidong Han, Hanjie Xu+10

KOKONI 3D, Moxin Technology · Zhejiang University · Huzhou University+4

3D Reconstruction 3D ViTs Multi-View 3D Reconstruction Feed-Forward 3D Reconstruction Efficient ViTs

Open paper AskPDF arXiv

High-resolution imagery is essential for accurate 3D reconstruction, as many geometric details only emerge at fine spatial scales. Recent feed-forward approaches, such as the Visual Geometry Grounded Transformer (VGGT), have demonstrated the ability to infer scene geometry from large collections of images in a single forward pass. However, scaling these models to high-resolution inputs remains challenging: the number of tokens in transformer architectures grows rapidly with both image resolution and the number of views, leading to prohibitive computational and memory costs. Moreover, we observe that visually ambiguous regions, such as repetitive patterns, weak textures, or specular surfaces, often produce unstable feature tokens that degrade geometric inference, especially at higher resolutions. We introduce HD-VGGT, a dual-branch architecture for efficient and robust high-resolution 3D reconstruction. A low-resolution branch predicts a coarse, globally consistent geometry, while a high-resolution branch refines details via a learned feature upsampling module. To handle unstable tokens, we propose Feature Modulation, which suppresses unreliable features early in the transformer. HD-VGGT leverages high-resolution images and supervision without full-resolution transformer costs, achieving state-of-the-art reconstruction quality.

Abstract

3D ViTs

Explore similar work

HD-VGGT: High-Resolution Visual Geometry Transformer

TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention

VGGT-Prime: Compute-Adaptive Mixture-of-Heads for Efficient Visual Geometry Transformers