WorksComputer Science
SGAR: Feed-Forward 3D Reconstruction via Sparse Geometry-Guided Attention
Abstract
Feed-forward models have been very successful in 3D reconstruction with ad hoc attention designs that relate all view patches densely. However, the cost of this dense global attention grows quadratically with the number of views, preventing the model from applying to larger-scale reconstruction settings. In this paper, we argue that most dense attention computations are unnecessary from a classical structure-from-motion perspective: 3D structures stem from geometrically matching regions in 2D. We bring this principle to learning-based feed-forward reconstruction with our Sparse Geometric Attention Reconstruction (SGAR) method. In SGAR, each patch token attends to only a small set of patch tokens from other views, chosen by a lightweight learned key selection module trained to predict cross-view 3D proximity of tokens. Thus, dense global attention is replaced by much more efficient geometric-similarity-guided sparse attention. SGAR is a general method that can be applied to different feed-forward models by fine-tuning. In our experiments, we show that SGAR is the fastest among both state-of-the-art dense and sparse methods and the most accurate sparse one. It runs up to 12-times faster than dense attention at 1,000 views with stable pose accuracy where existing sparse methods fail, and shows the best performance in reconstructing more complete geometries on the same time budget.
Introduction
Multi-view scene reconstruction is a long-standing problem in computer vision. Classical Structure-from-Motion (SfM) solutions treat this as a geometric estimation problem and solve it with a multi-stage pipeline built on projective geometry and robust optimization. Although effective, SfM methods are based on complex pipelines and produce only sparse reconstructions [34, 36, 37, 38, 46, 53]. Recent feed-forward models such as VGGT [60] and DejaView [8] replace it by a single network that predicts cameras and geometry for all views in a single forward pass, matching or even surpassing the earlier methods in a fraction of the time when the views overlap substantially. At the core of these methods is global attention: The tokens of all views are processed jointly, so that every view is interpreted in the context of every other. This is what makes these approaches so powerful, but also what makes them expensive as the cost of global attention grows quadratically with the total number of views. For example, in DejaView [8], the time cost grows from a few seconds in 100 views to more than 300 s in 1,000 views on one Nvidia A100 GPU. This severely restricts the applicability of feed-forward models to time-sensitive applications, such as robot navigation, and to providing a fast initialization for optimization-based reconstruction pipelines.
However, much of this computation is redundant. As shown in Figure 2, global attention is highly sparse even in densely trained models: When each query of DejaView attends only to a small fraction of the oracle keys determined based on ground truth geometry, the model retains most of the pose accuracy without any retraining. Many recent works leverage this observation to accelerate VGGT-style models, mostly without retraining. The strategies include merging similar tokens [11, 48], compressing each view into a few descriptors [66], test-time training [20, 23], or estimating the attention map at a coarse resolution and keeping only its strongest blocks [16, 41, 52, 58]. These methods infer the relevant keys from an approximation of the attention itself, rather than from the geometric relations between views on which reconstruction depends. As a consequence, their accuracy degrades significantly compared to dense methods, or they run out of memory on long sequence.
Figure 2. Dense attention is redundant. By selecting oracle top- keys based on ground-truth geometry, most of the pose accuracy can be retained even when 99% of attention weights are zeroed out. Evaluating DejaView [8] on an NRGBD [3] scene.
In this work, we revisit this problem from the perspective of classical Structure-from-Motion, in which views are related through sparse 2D correspondences. We hypothesize that, apart from a few tokens that anchor the global coordinate frame, the keys a patch token requires are the tokens that observe the same 3D surface region in other views. Key selection then becomes a geometric-feature retrieval problem, which can be supervised directly with ground-truth geometry and solved without computing dense attention.
Building on this insight, we propose SGAR, which replaces the dense global attention of feed-forward reconstruction models with geometry-guided sparse attention (Figure 1). A lightweight projection, trained with a contrastive loss on ground-truth geometry, maps patch tokens into an embedding space in which tokens observing the same surface lie close together. At inference, these embeddings are clustered, and each patch token attends only to the patches of its nearest clusters, together with a small set of tokens that provide global context. Since the number of keys per token grows much more slowly than the total number of tokens, the cost of global attention becomes sub-quadratic in the number of views. SGAR can be applied to different dense attention models such as DejaView [8] and VGGT- [61] by fine-tuning pretrained weights and requires no ground-truth geometry at inference.
Figure 1. SGAR replaces the dense global attention of feed-forward reconstruction with a sparse one: besides a few special tokens, each token attends only to the tokens that the model predicts to lie near it in 3D, which makes inference fast. Three tokens of the query view are marked in color, and the other views show the patches the model predicts for them. On this scene with 1,000 views, SGAR takes 26.1 s against 319.6 s for DejaView on one Nvidia A100 GPU, at an AUC@30 of 84.6 against 86.5.
Experiments on six datasets show that SGAR is the fastest among state-of-the-art dense and sparse methods, running up to faster than dense models at 1,000 views and keeping accuracy comparable to dense models in all scenarios. Because of its high efficiency, we also show that SGAR can handle many more images and cover much more extended scenes for a given computational budget.
In summary, our contributions are as follows.
We show that global attention in feed-forward reconstruction is highly sparse, and that the keys a query token requires can be predicted by their geometric similarities.
We propose SGAR, a geometry-guided sparse attention mechanism whose keys are selected by a contrastively learned embedding, obtained by fine-tuning a pretrained dense model.
We demonstrate that SGAR achieves the best trade-off between efficiency and accuracy, scaling to 1,000 views on a single GPU, and reconstructing the most complete geometries at the same given time budget.
Code and models will be released.
Related Work
Feed-forward 3D reconstruction. Structure-from-Motion pipelines [34, 36, 38, 46, 53] estimate poses and structure in stages, from keypoint matching to bundle adjustment, and learned matchers [30, 44, 51] have made each stage more robust. DUSt3R [63] and MASt3R [27] showed that a network can instead predict aligned pointmaps directly, but they operate on two images at a time and need an optimization to merge the pairwise outputs. VGGT [60] and the models that followed it [29, 61, 65, 70] read an arbitrary number of images at once, interleaving attention within each image and across all of them, and output cameras, depth and points without any optimization. DejaView [8] reaches comparable accuracy with far fewer parameters by applying one shared block recurrently. Streaming models [59, 62] take a different route and ingest the images one by one, updating a persistent memory. For the models that process all views jointly, the cost of dense global attention grows quadratically with the number of views. However, we observe that dense attention is highly redundant (Figure 2), and SGAR therefore restricts attention to sparse, geometrically similar tokens.
Sparse attention. A growing body of work accelerates VGGT-style models, most of it without retraining. Token merging methods, following token merging for vision transformers [7], reduce the tokens that enter global attention, merging near-duplicate patches [48] or low-confidence ones ranked by a distilled predictor [11]. FlashVGGT [66] compresses each view into a few descriptor tokens and turns global attention into cross-attention to them. VGG-T³ [20] and ZipMap [23] distill the key-value set into a fixed-size network by test-time training. AVGGT [52] turns the early global layers into frame attention and subsamples the keys of the remaining ones on the patch grid, and Lite3R [74] distills a sparse linear attention into the frozen backbone. Block-Sparse [58] pools the keys into blocks and reads only those that hold the attention mass, SAF3R [16] fits a sparse pattern to each head, and Speed3R [41] pairs a compressed branch with a sparse one that attends to the blocks it selects. Most of these methods approximate dense attention at the granularity of blocks or groups of tokens at the risk of oversimplification. SGAR instead selects the exact keys that are meaningful for every token individually from a geometric point of view.
Restricting attention to a subset of keys is a long-standing remedy in the broader literature too. Early methods fix the pattern in advance, with local windows and strided or global tokens [5, 12, 31, 72], while content-based methods group queries and keys by hashing [25] or clustering [43], or retrieve the nearest keys from a large memory [68]. We investigate the problem in a different context of geometric reconstruction and learn key selection by training geometric representations.
Learned geometric embeddings. Finding correspondences between images rests on learning geometric embeddings of points whose similarity reflects whether two locations observe the same 3D point. Learned geometric embeddings are trained with metric or contrastive objectives on ground-truth correspondences [13]. Modern matchers reason over both images with attention, on sparse key-points [30, 44, 55] or densely [19, 51]. MASt3R [27] brings matching into feed-forward reconstruction with a head trained by an InfoNCE loss [35], whose descriptors are matched after the forward pass. Supervised contrastive learning [24] extends InfoNCE to several positives per anchor, as when a patch overlaps several patches of another view. SGAR uses the same idea as a supervision signal during training: a multi-positive InfoNCE loss pulls together the embeddings of corresponding patches and pushes away all other patches. However, our learning objective is not to find point correspondences, but to learn patch-level geometric embeddings that are never matched explicitly. The geometric similarities will be used to sparsify global attention with reconstruction as the final objective.
Method
We introduce SGAR, a sparse-attention reconstruction scheme in which attention across views is restricted to patches that correspond to the same 3D neighborhood. This makes its cost sub-quadratic in the number of views as opposed to quadratic in dense attention. It rests on two things: a learned key selection that clusters the tokens from different views likely to be in 3D proximity, and a sparse global attention mechanism in which every patch token attends in full only to the patches of its own clusters and the 3 nearest clusters, as well as to the rest of the scene through a few global tokens and cluster summaries. SGAR can be readily applied to any dense attention models such as DejaView [8] and VGGT-Ω [61], by replacing its global attention with our sparse ones and fine-tuning. This yields the complete architecture of Figure 3, where DejaView is used as an example base architecture. We define the key selection process in § 3.1, describe our sparse attention mechanism in § 3.2, and present our training procedure in § 3.3.

Figure 3. Overview of SGAR. We replace the global dense attention of DejaView [8] with our sparse geometric attention as an example. The images are encoded by DINOv2 and augmented with camera and register tokens. A single shared block, unrolled over its steps, runs frame attention, key selection and sparse attention. In sparse attention, each patch attends to the special tokens of all views, the patches of the reference view, a summary of each cluster, and all the patches of the nearest clusters, so that its cost grows near-linearly with the number of views. Key Selection projects the patch tokens by into a geometric embedding and clusters the embeddings by spherical -means into clusters, re-clustered at every step. Each cluster probes its nearest clusters, itself first. With the patches sorted by cluster, the selected keys form a block-sparse mask over tiles of 128 tokens. A contrastive loss pulls together the embeddings of patches that observe the same surface in the ground-truth geometry and pushes the others apart.
Notation. The input is a set of RGB images of the same scene . Each image is patchified into tokens and augmented with one camera token and four register tokens [15], which we jointly call the special tokens, giving tokens of dimension per view. The tokens are refined by a number of alternations of frame attention and global attention. We write for the number of patch tokens, index a patch token by , and write for the special tokens and for the patch tokens of the first view, which is the reference view in whose camera frame the model predicts the scene.
Learning key selection
Key selection aims to identify the relevant patches that will restrict attention in § 3.2. We first compute geometric embeddings for every patch token with a learned projection , where and are close in feature space if and only if and belong to the same neighborhood in 3D. We then consider these embeddings across all views, and cluster them in feature space. By design, the resulting clusters correspond to close 3D regions and will be used to restrict attention computation in § 3.2.
Geometric embedding. The shared block carries a linear adjacency projection , applied to the patch tokens after frame attention, as is common in contrastive learning [10, 39]. Its outputs are centered over the patches of the sample and normalized,
so that two patches are compared by the cosine of their embeddings, which is a normalized Mahalanobis inner product with a metric learned as in [67]. Centering removes the component shared by all patches of a scene, which would otherwise dominate the cosines and pull every embedding into one direction. The embedding is trained so that patches observing the same surface point have nearby embeddings, with the contrastive loss defined below.
Contrastive loss. We train with a multi-positive InfoNCE [24, 35] on the same cosine that the -means clusters by. We sample a set of 1024 query patches with valid ground truth and keep those with at least one corresponding patch. For a query , let be its corresponding patches and the candidates: all patches of the other views, except those within five patches of a corresponding patch that do not correspond themselves, whose status is an artefact of the patch grid. With the embedding of Eq. (1) at layer , the loss is
averaged over the samples of the batch, where is a learned temperature initialised at 20 and capped at 100. The projection reads the tokens through a stop-gradient, so trains and alone, while the reconstruction losses train the backbone. The clustering itself is not differentiable, the backbone learns to use the clusters it is given.
To generate the ground truth, we call two patches of different views corresponding when they observe the same surface. Patch of view corresponds to patch of view if a pixel of , unprojected with its ground-truth depth and reprojected into , lands within one pixel of a pixel of whose ground-truth depth matches to within 5%.
Key selection by clustering. Instead of ranking the keys of every query individually, we group all patches at once. The embeddings are clustered by spherical -means [17], seeded by -means++ [1], into clusters: every patch is assigned to the centroid with the highest cosine similarity, and every centroid is set to the normalized mean of its patches. Each cluster then selects the clusters whose centroids are most similar to its own, itself first. A patch of cluster attends to all patches of these clusters,
The nearest clusters cover the patches at the boundary of a cluster, and the surfaces that -means splits into several clusters, which a single cluster would miss. The selection is made per cluster rather than per patch, so that all queries of a cluster share their keys; this is what lets the attention run as dense tiles on tensor cores (see § 3.2). The clustering is redone at every layer, so that it follows the model’s evolving estimate of the scene.
Sparse global attention
With the learned key selection, dense attention can be sparsified. We replace a dense model’s global attention with a few residual updates that reuse the model’s global attention weights. First, we let the special tokens attend densely to all tokens to preserve a global channel that the clusters do not restrict. Each patch token is then updated by four attentions computed from one projection of the patch tokens, each with its own softmax, whose outputs are summed and passed through the output projection :
where is the head dimension, and the four key sets in are the special tokens , the patches of the reference view, the patches of the nearest clusters, and one summary token for every cluster that is not in .
Implementation and complexity. The patches are sorted by cluster, and every cluster is arranged to contain contiguous blocks that are padded to 128, the tile of the attention kernel. Every tile of 128 queries then belongs to a single cluster and attends to the whole tiles of closest clusters, so the attention over is a block-sparse FlexAttention [18] whose mask we build directly from the block boundaries. A patch attends to keys on average, special tokens, reference patches, summaries and four clusters of patches on average, instead of . Global attention thus costs rather than , and each round of -means costs . At 1,000 views of 972 patches, a patch attends to about 11,000 keys, about 1% of the tokens.
Training
We train SGAR by fine-tuning from a pretrained dense attention model (e.g. DejaView, VGGT-). The adjacency projection and the temperature of are the only new parameters. The model is trained with the following loss:
where are the dense model’s point-map, depth, ray and camera losses on the ground truth, expressed in the camera frame of the reference view after scale normalisation. The second term is the contrastive loss (2), with .
Training Details. We train on a mixture of 18 public datasets that covers indoor and outdoor scenes, driving, aerial and object-centric captures, both real and synthetic: ScanNet [14], Hypersim [42], Replica [50], Taskonomy [73], CA-1M [26], Aria Synthetic Environments [2], MegaDepth [28], BlendedMVS [71], CO3D [40], renderings of TRELLIS 3D assets [69], Mapillary Metropolis [32], Virtual KITTI 2 [9], TartanAir and its V2 release [64], MVS-Synth [22], UnrealStereo4K [54], Spring [33] and Kubric-4D [21, 57]. A sample of views is drawn only from scenes with at least frames, without repeated frames. In training runs, the view samples hold 2 to 32 views. The aspect ratio of images is drawn per batch between 1:3 and 1:1, and receives color, blur, and zoom augmentation. Since the dense model’s representation is to be fine-tuned rather than relearned, the pretrained weights are trained at a peak learning rate of and the new parameters at , with AdamW. We train on 4 H100 GPUs in bfloat16, with gradient checkpointing of every layer, and evaluate the checkpoint after 50k steps.
Results
Experimental setup
Benchmark. We evaluate zero-shot camera pose and depth estimation in three ranges of sequence length, on five public datasets that are not in our training data, 200 scenes in total, and on one long capture from the ScanNet test split, which is held out from training. Short: 7-Scenes [49], ETH3D [47], LaMAR [45] and NRGBD [3], each scene subsampled at a fixed frame rate and evaluated on its first views. Mid-range: the 9 scenes of Mip-NeRF 360 [4], each on all of its 125 to 311 views (194 median); they are photo collections around an object rather than videos, so the views are far apart but all observe the same object. Long: ScanNet [14] test scene0770_00, a 3,383-frame capture whose camera covers the largest floor area of the test set, swept from to 1,000 views spaced evenly over the capture.
Metrics. Following [60], the error of a view pair is the larger of its relative rotation and translation angular errors, and we report the area under the pair-error curve up to (AUC@30) and (AUC@5), averaged over the scenes of a dataset and then over datasets. Depth is evaluated on the short datasets with ground-truth depth, 7-Scenes, LaMAR and NRGBD: the predicted depth is aligned to the ground truth with one median scale per sequence, so that it must be consistent across views, and we report the share of pixels with .
Baselines. We apply SGAR to two pretrained dense models of very different size and design, and evaluate both: SGAR (DejaView), fine-tuned from DejaView [8], and SGAR (VGGT-), fine-tuned from VGGT- [61]. compare with the dense VGGT [60], [65], Depth Anything 3 (DA3, giant) [29], VGGT- [61] and DejaView [8]. We also compare with ten efficient models: Block-Sparse [58], FastVGGT [48], SAF3R [16], AVGGT [52] at two sparsity levels , FlashVGGT [66], Co-Me [11], the causal streaming StreamVGGT [75], and VGG-T [20] and ZipMap [23] which replaces global attention with test-time training layers of linear cost. Each model runs with its released weights. All models run on one A100 80 GB in bfloat16, and we time the full forward pass of a scene, excluding image loading and a warm-up run per input shape.
Comparison with the state of the art
Accuracy and cost. Table 1 compares all models in the three ranges of sequence length. Both variants are the fastest models in every range, and in every range they are also the most accurate of the efficient models. SGAR also greatly improves the efficiency of the pretrained dense attention models while maintaining comparable accuracies. On the long ScanNet sequence, SGAR runs up to faster than the dense models at close pose accuracy, whereas every other efficient model runs out of memory, fails, or loses most of its accuracy. We also compare the depth estimation results in Table 3, our SGAR (VGGT-) is the most accurate efficient model, which matches the performance of VGGT-. We also provide qualitative results in Figure 4 with comparison to FastVGGT and ZipMap, where our method shows both the best efficiency and point map quality.
| Method | Params | 7-Scenes | ETH3D | LaMAR | NRGBD | @5 | @30 | Time | @5 | @30 | Time | @5 | @30 | Time | Mem |
| VGGT [60] | 1158 | 81.0 | 70.6 | 27.3 | 97.2 | 40.2 | 69.0 | 6.98 | 74.5 | 94.9 | 54.1 | OOM | – | – | 75.3 |
| X [65] | 1360 | 82.0 | 83.7 | 59.1 | 98.6 | 51.8 | 80.8 | 3.82 | OOM | – | – | OOM | – | – | – |
| DA3 [29] | 1356 | 82.8 | 83.6 | 47.7 | 99.8 | 53.5 | 78.5 | 3.81 | 90.9 | 98.3 | 24.3 | OOM | – | – | 39.4 |
| VGGT- [61] | 1144 | 88.2 | 90.1 | 59.7 | 98.6 | 57.9 | 84.2 | 2.27 | 86.3 | 97.4 | 14.2 | 27.9 | 85.3 | 303.8 | 22.9 |
| DejaView [8] | 117 | 83.7 | 87.8 | 59.2 | 98.6 | 48.6 | 82.3 | 2.10 | 85.6 | 97.5 | 14.1 | 31.1 | 86.5 | 319.6 | 27.4 |
| Block-Sparse [58] | 1158 | 80.8 | 70.5 | 27.2 | 97.1 | 39.8 | 68.9 | 4.31 | 73.3 | 94.7 | 28.9 | OOM | – | – | 26.6 |
| FastVGGT [48] | 1158 | 79.5 | 66.4 | 26.6 | 96.2 | 35.3 | 67.2 | 3.47 | 72.3 | 94.5 | 16.5 | 0.0 | 3.2 | 257.0 | 26.3 |
| SAF3R [16] | 1256 | 79.7 | 56.3 | 21.1 | 95.6 | 32.0 | 63.2 | 3.55 | 74.3 | 94.9 | 18.0 | FAIL | – | – | 26.9 |
| AVGGT [52], | 1158 | 80.2 | 63.3 | 21.4 | 96.7 | 37.3 | 65.4 | 3.52 | 72.2 | 93.8 | 20.6 | OOM | – | – | 76.4 |
| AVGGT [52], | 1158 | 79.6 | 59.9 | 20.7 | 96.4 | 35.6 | 64.1 | 2.76 | 70.4 | 93.2 | 13.6 | OOM | – | – | 76.0 |
| FlashVGGT [66] | 1158 | 77.6 | 57.5 | 16.7 | 93.1 | 29.1 | 61.2 | 2.12 | 58.8 | 90.2 | 8.1 | OOM | – | – | 28.0 |
| Co-Me [11] | 1191 | 75.4 | 56.9 | 9.8 | 87.5 | 21.8 | 57.4 | 2.09 | 36.6 | 81.4 | 11.3 | 3.9 | 35.4 | 203.4 | 23.0 |
| StreamVGGT [75] | 1256 | 75.7 | 54.7 | 17.6 | 88.9 | 25.9 | 59.2 | 13.95 | 39.1 | 79.3 | 109.6 | OOM | – | – | 42.7 |
| VGG-T [20] | 1192 | 68.1 | 44.0 | 14.1 | 88.2 | 20.1 | 53.6 | 3.80 | 15.2 | 65.1 | 11.1 | OOM | – | – | 49.0 |
| ZipMap [23] | 1334 | 80.5 | 73.4 | 39.6 | 96.5 | 39.8 | 72.5 | 1.81 | 61.6 | 90.8 | 5.5 | OOM | – | – | 32.3 |
| SGAR (VGGT-) | 1164 | 84.0 | 88.6 | 54.5 | 97.7 | 51.3 | 81.2 | 1.65 | 81.9 | 96.5 | 5.4 | 23.5 | 83.5 | 40.5 | 23.7 |
| SGAR (DejaView) | 118 | 82.3 | 83.6 | 49.7 | 97.3 | 42.7 | 78.2 | 1.32 | 78.3 | 96.0 | 4.1 | 27.1 | 84.6 | 26.1 | 28.5 |
Table 1. Camera pose estimation. SGAR (VGGT-) and SGAR (DejaView) are the fastest models at every sequence length and the most accurate efficient ones, close to their dense counterparts. AUC in %, time in s per scene, Params in millions, Mem in peak GiB on Mip-NeRF 360; OOM: out of memory; FAIL: kernel compile error; –: not available. Green, yellow-green, yellow: first, second, third; ties share a place.
| Method | 7-Scenes | LaMAR | NRGBD | Mean |
| VGGT [60] | 95.4 | 81.1 | 99.5 | 92.0 |
| X [65] | 98.5 | 94.0 | 99.6 | 97.3 |
| DA3 [29] | 97.5 | 91.6 | 99.5 | 96.2 |
| VGGT- [61] | 97.4 | 91.8 | 100.0 | 96.4 |
| DejaView [8] | 97.6 | 87.3 | 99.3 | 94.8 |
| Block-Sparse [58] | 95.4 | 81.0 | 99.5 | 92.0 |
| FastVGGT [48] | 95.8 | 81.0 | 99.4 | 92.1 |
| SAF3R [16] | 95.8 | 74.9 | 99.1 | 89.9 |
| AVGGT [52], | 95.5 | 74.9 | 99.4 | 90.0 |
| AVGGT [52], | 95.7 | 73.9 | 99.5 | 89.7 |
| FlashVGGT [66] | 96.1 | 72.9 | 99.5 | 89.5 |
| Co-Me [11] | 94.0 | 73.9 | 95.7 | 87.9 |
| StreamVGGT [75] | 94.7 | 60.3 | 99.1 | 84.7 |
| VGG-T [20] | 96.6 | 89.6 | 99.5 | 95.2 |
| ZipMap [23] | 95.3 | 93.0 | 99.7 | 96.0 |
| SGAR (VGGT-) | 98.4 | 91.1 | 99.9 | 96.5 |
| SGAR (DejaView) | 98.1 | 76.5 | 99.3 | 91.3 |
Table 3. Depth estimation. Depth on the short benchmark ( views per scene) for the three datasets with ground-truth depth. The predicted depth is aligned to the ground truth with one median scale per sequence, so it must be consistent across views. : share of pixels with , in per cent (higher is better). Mean is over the three datasets, each weighted equally. Of each column, over all models, green marks the best value, yellow-green the second and yellow the third; tied values share a place.

Figure 4. Qualitative comparison. Point clouds predicted by FastVGGT, ZipMap and SGAR (DejaView), with the ground truth, on one scene of each of 7-Scenes (chess/seq-03), ETH3D (courtyard, laser scan), LaMAR (HGE) and Mip-NeRF 360 (bonsai, SfM points). Left: four of the input views. Below each reconstruction: the AUC@30 of its cameras. SGAR (DejaView) keeps the structure of the scene where the other efficient models fragment it, most visibly on the long LaMAR trajectory.
Scaling with the number of views. Figure 5 provides testing results on the long ScanNet capture up to 1,000 views. The time of both our variants grows far more slowly than that of the dense models, and their accuracy stays nearly flat. At 1,000 views, SGAR (DejaView) is faster than DejaView and VGGT-, while maintaining comparable accuracies. SGAR (VGGT-) is 7.5× faster than VGGT- at 1,000 views and retaining most of its AUC@30. The other efficient models run out of memory or fail before 1,000 views, or lose much of their accuracy.
Figure 5. Scaling with the number of views. Time per scene (a, log scale) and AUC@30 (b) on the ScanNet test capture scene0770_00, with 100 to 1,000 views evenly spaced over its 3,383 frames, so that more views sample the same trajectory more densely. A line ends where the model runs out of memory on the A100 80 GB or fails; the cross marks its last size.
Reconstruction at a fixed time budget. To measure what the speed buys, we also fix the inference time instead of the number of views for a time budget experiment as in Table 2. We use the ScanNet capture of the long benchmark and budgets from 1 to 20 s. For each model and budget, we find the largest number of views, spread evenly over the capture, whose forward pass fits in the budget. Every pixel’s predicted depth is unprojected with the predicted cameras, aligned to the ground truth by the similarity that maps the predicted cameras onto the ground-truth ones, refined by a similarity ICP [6, 56] onto the ground-truth surface, and voxelised at 1 cm. We report the reconstruction coverage @5 cm, i.e. the share of the ground-truth mesh within 5 cm of a predicted point. Because of their efficiency, SGAR (DejaView) and SGAR (VGGT-) both fit many more views than the corresponding dense models into the same time budget, and these extra views translate into a more complete reconstruction. Moreover, our method also compares favorably against all sparse models.
| Method | 1 | 2 | 3 | 5 | 7.5 | 10 | 15 | 20 |
| VGGT [60] | 59.5 | 69.9 | 73.2 | 78.6 | 81.2 | 84.4 | 87.0 | 88.3 |
| X [65] | 55.9 | 67.5 | 73.5 | 80.1 | 80.6 | 81.9 | 84.9 | 85.5 |
| DA3 [29] | 59.8 | 74.4 | 75.9 | 83.6 | 84.7 | 87.3 | 88.8 | 89.9 |
| VGGT- [61] | 71.0 | 79.6 | 82.5 | 86.4 | 88.7 | 90.7 | 90.9 | 92.4 |
| DejaView [8] | 73.8 | 78.6 | 84.2 | 86.2 | 87.0 | 88.2 | 89.2 | 91.3 |
| Block-Sparse [58] | 60.4 | 73.1 | 77.9 | 82.7 | 85.2 | 87.0 | 89.1 | 90.4 |
| FastVGGT [48] | 45.1 | 71.2 | 79.9 | 84.9 | 87.4 | 89.2 | 90.0 | 91.8 |
| SAF3R [16] | 11.6 | 71.2 | 80.9 | 87.7 | 89.0 | 90.9 | 92.2 | 93.2 |
| AVGGT [52], | 59.5 | 75.9 | 79.5 | 84.5 | 86.4 | 89.6 | 91.1 | 92.1 |
| AVGGT [52], | 64.5 | 79.2 | 81.5 | 87.0 | 87.5 | 90.2 | 91.6 | 93.6 |
| FlashVGGT [66] | 55.7 | 78.0 | 84.3 | 88.9 | 90.7 | 91.2 | 92.9 | 93.4 |
| Co-Me [11] | 59.0 | 73.1 | 80.9 | 83.4 | 86.8 | 87.3 | 90.2 | 91.4 |
| StreamVGGT [75] | 11.6 | 11.2 | 39.1 | 57.8 | 63.5 | 69.5 | 74.2 | 73.1 |
| VGG-T [20] | 9.2 | 60.5 | 62.3 | 70.4 | 84.3 | 82.2 | 79.2 | 74.2 |
| ZipMap [23] | 65.0 | 79.8 | 85.2 | 88.1 | 89.3 | 90.1 | 92.2 | 92.9 |
| SGAR (VGGT-) | 80.7 | 89.4 | 90.6 | 93.3 | 94.0 | 95.2 | 95.5 | 95.9 |
| SGAR (DejaView) | 77.3 | 84.1 | 88.1 | 90.8 | 92.5 | 92.8 | 93.8 | 94.1 |
Table 2. Reconstruction at a fixed time budget. Coverage @5 cm (%): share of the ground-truth surface of ScanNet scene0770_00 within 5 cm of a predicted point after ICP alignment, with each model given as many views as fit in the budget. Of each column, over all models, green marks the best value, yellow-green the second and yellow the third. Our model can recover the most complete geometries at all time budgets.
Ablation study
In Table 4, we provide an ablation study regarding different aspects of our method. First, we investigate which keys a patch attends to. Attending to every patch of the nearest clusters is the most accurate choice. Keeping only the 256 of them with the highest adjacency scores, as a per-query top- retrieval would, loses significant accuracy of AUC@30 in every setting. It is also slower, since selecting the keys costs more than attending to fewer of them saves.
| Variant | @30 | Time | Mip360 @30 | Mip360 Time | ScanNet @30 | ScanNet Time |
| DejaView, dense | 82.3 | 2.10 | 97.5 | 14.1 | 86.5 | 319.6 |
| top-256 | 68.8 | 1.85 | 88.4 | 5.6 | 75.1 | 33.6 |
| -means on raw tokens | 65.3 | 1.39 | 86.7 | 4.4 | 71.9 | 30.3 |
| w/o reference view | 74.1 | 1.30 | 93.1 | 4.0 | 84.7 | 25.6 |
| w/o summaries | 76.3 | 1.29 | 95.0 | 3.9 | 83.5 | 24.3 |
| w/o and summaries | 65.1 | 1.26 | 86.4 | 3.8 | 79.2 | 23.8 |
| SGAR (DejaView) | 78.2 | 1.32 | 96.0 | 4.1 | 84.6 | 26.1 |
Table 4. Ablations. An ablation study on the SGAR (DejaView) model. AUC@30 (%) and time per scene. : mean over the four short datasets; ScanNet: 1,000 views.
Second, we investigate what the clusters should be built on. Running -means on the raw tokens instead of the learned embeddings is the largest single loss. This means that we need a good geometric representation of tokens, and the contrastive projection is what finds them. Third, we investigate the effectiveness of coarse key sets. Without the reference view, AUC@30 drops significantly at views, and it also drops noticeably without the cluster summaries. Without both, it drops far more than the sum, to the level of raw-token clustering. This is because each set gives a patch a path to the scene beyond its nearest clusters, and at least one is needed. Together they add only keys per patch and only 5 to 9% of the time.
Conclusion and Discussion
In this paper, we presented SGAR, a general method that replaces the dense global attention of feed-forward reconstruction models with geometry-guided sparse attention. A learned projection embeds each patch such that patches observing the same surface lie close together, and our sparse geometric attention then selects a small subset of relevant keys for each query by clustering in this embedding space. We applied SGAR to two state-of-the-art dense-attention models and showed that it is the fastest model at every sequence length, up to 12× faster than dense attention at 1,000 views, and the most accurate of the efficient alternatives, with accuracy comparable to the original dense models. Despite the good performance, SGAR also comes with limitations: sparsification still brings a small drop in accuracy relative to dense attention. Closing this gap is a promising direction for future work.
References
- [1]David Arthur and Sergei Vassilvitskii. k-means++: The advantages of careful seeding. In ACM-SIAM Symposium on Discrete Algorithms, 2007.DOI
- [2]Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, et al. Scene-Script: Reconstructing scenes with an autoregressive structured language model. In European Conference on Computer Vision, pages 247–263, 2024.arxiv.org/abs/2403.13064
- [3]Dejan Azinović, Ricardo Martin-Brualla, Dan B. Goldman, Matthias Nießner, and Justus Thies. Neural RGB-D surface reconstruction. In Conference on Computer Vision and Pattern Recognition, 2022.arxiv.org/abs/2104.04532
- [4]J. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman. Mip-NeRF 360: Unbounded Anti-Aliased Neural Radiance Fields. In Conference on Computer Vision and Pattern Recognition, 2022.arxiv.org/abs/2111.12077
- [5]Iz Beltagy, Matthew E. Peters, and Arman Cohan. Longformer: The long-document transformer. arXiv Preprint, 2004.05150, 2020.arxiv.org/abs/2004.05150
- [6]P. Besl and N. Mckay. A Method for Registration of 3D Shapes. IEEE Transactions on Pattern Analysis and Machine Intelligence, 14(2):239–256, 1992.
- [7]Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your ViT but faster. In International Conference on Learning Representations, 2023.arxiv.org/abs/2210.09461
- [8]Alessandro Burzio, Tobias Fischer, Sven Elflein, Qunjie Zhou, Riccardo de Lutio, Jiawei Ren, Jiahui Huang, Shengyu Huang, Marc Pollefeys, Laura Leal-Taixé, Zan Gojcic, and Haithem Turkı̇. Déjà view: Looping transformers for multi-view 3d reconstruction. arXiv Preprint, 2605.30215, 2026.arxiv.org/abs/2605.30215
- [9]Yohann Cabon, Naila Murray, and Martin Humenberger. Virtual KITTI 2. arXiv preprint arXiv:2001.10773, 2020.
- [10]T. Chen, S. Kornblith, M. Norouzi, and G. Hinton. A Simple Framework for Contrastive Learning of Visual Representations. In International Conference on Machine Learning, 2020.arxiv.org/abs/2002.05709
- [11]Yutian Chen, Yuheng Qiu, Ruogu Li, Jay Patrikar, and Sebastian Scherer. Co-me: Confidence guided token merging for visual geometric transformers. In Conference on Computer Vision and Pattern Recognition, pages 14590–14599, 2026.arxiv.org/abs/2511.14751
- [12]Rewon Child, Scott Gray, Alec Radford, and Ilya Sutskever. Generating long sequences with sparse transformers. arXiv Preprint, 1904.10509, 2019.arxiv.org/abs/1904.10509
- [13]C. Choy, J. Gwak, S. Savarese, and M. Chandraker. Universal Correspondence Network. In Advances in Neural Information Processing Systems, 2016.arxiv.org/abs/1606.03558
- [14]A. Dai, A.X. Chang, M. Savva, M. Halber, T. Funkhouser, and M. Nießner. Scannet: Richly-Annotated 3D Reconstructions of Indoor Scenes. In Conference on Computer Vision and Pattern Recognition, 2017.arxiv.org/abs/1702.04405
- [15]Timothée Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. In International Conference on Learning Representations, 2024.arxiv.org/abs/2309.16588
- [16]Jianing Deng, Yuanzhe Li, Jialu Wang, Song Wang, Tianlong Chen, Huanrui Yang, and Jingtong Hu. Saf3r: Dynamic sparse attention for feed-forward 3d reconstruction transformers. arXiv Preprint, 2607.03612, 2026.arxiv.org/abs/2607.03612
- [17]I. S. Dhillon and D. S. Modha. Concept Decompositions for Large Sparse Text Data Using Clustering. Machine Learning, 42(1):143–175, 2001.DOI
- [18]J. Dong, B. Feng, D. Guessous, Y. Liang, and H. He. Flex Attention: A Programming Model for Generating Optimized Attention Kernels. In Conference on Machine Learning and Systems, 2025.arxiv.org/abs/2412.05496
- [19]Johan Edstedt, Qiyu Sun, Georg Bökman, Mårten Wadenbäck, and Michael Felsberg. Roma: Robust dense feature matching. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19790–19800. IEEE, 2024.DOI
- [20]Sven Elflein, Ruilong Li, Sérgio Agostinho, Zan Gojcic, Laura Leal-Taixé, Qunjie Zhou, and Aljosa Osep. Vggt-3: Offline feed-forward 3d reconstruction at scale. In Conference on Computer Vision and Pattern Recognition, 2026.arxiv.org/abs/2602.23361
- [21]Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In Conference on Computer Vision and Pattern Recognition, pages 3749–3761, 2022.arxiv.org/abs/2203.03570
- [22]P. Huang, K. Matzen, J. Kopf, N. Ahuja, and J. Huang. Deepmvs: Learning Multi-View Stereopsis. In Conference on Computer Vision and Pattern Recognition, pages 2821–2830, 2018.arxiv.org/abs/1804.00650
- [23]Haian Jin, Rundi Wu, Tianyuan Zhang, Ruiqi Gao, Jonathan T. Barron, Noah Snavely, and Aleksander Holynski. ZipMap: Linear-time stateful 3D reconstruction with test-time training. In Conference on Computer Vision and Pattern Recognition, 2026.arxiv.org/abs/2603.04385
- [24]Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. In Advances in Neural Information Processing Systems, 2020.arxiv.org/abs/2004.11362
- [25]Nikita Kitaev, Łukasz Kaiser, and Anselm Levskaya. Reformer: The efficient transformer. In International Conference on Learning Representations, 2020.arxiv.org/abs/2001.04451
- [26]Justin Lazarow, David Griffiths, Gefen Kohavi, Francisco Crespo, and Afshin Dehghan. Cubify Anything: Scaling indoor 3D object detection. In Conference on Computer Vision and Pattern Recognition, pages 22225–22233, 2025.arxiv.org/abs/2412.04458
- [27]Vincent Leroy, Yohann Cabon, and Jérôme Revaud. Grounding image matching in 3d with MASt3R. In European Conference on Computer Vision, 2024.arxiv.org/abs/2406.09756
- [28]Zhengqi Li and Noah Snavely. Megadepth: Learning single-view depth prediction from internet photos. In Conference on Computer Vision and Pattern Recognition, 2018.arxiv.org/abs/1804.00607
- [29]Haotong Lin, Sili Chen, Jun Hao Liew, Donny Y. Chen, Zhenyu Li, Yang Zhao, Sida Peng, Hengkai Guo, Xiaowei Zhou, Guang Shi, Jiashi Feng, and Bingyi Kang. Depth anything 3: Recovering the visual space from any views. In International Conference on Learning Representations, 2026.arxiv.org/abs/2511.10647
- [30]Philipp Lindenberger, Paul-Edouard Sarlin, and Marc Pollefeys. LightGlue: Local feature matching at light speed. In International Conference on Computer Vision, 2023.arxiv.org/abs/2306.13643
- [31]Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In International Conference on Computer Vision, 2021.arxiv.org/abs/2103.14030
- [32]Mapillary. Mapillary Metropolis. https://www.mapillary.com/dataset/metropolis, 2021.
- [33]Lukas Mehl, Jenny Schmalfuss, Azin Jahedi, Yaroslava Nali-vayko, and Andrés Bruhn. Spring: A high-resolution high-detail dataset and benchmark for scene flow, optical flow and stereo. In Conference on Computer Vision and Pattern Recognition, pages 4981–4991, 2023.
- [34]Pierre Moulon, Pascal Monasse, Romuald Perrot, and Renaud Marlet. OpenMVG: Open multiple view geometry. In International Workshop on Reproducible Research in Pattern Recognition, pages 60–74. Springer, 2016.
- [35]Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representation Learning with Contrastive Predictive Coding. In arXiv Preprint, 2018.arxiv.org/abs/1807.03748
- [36]OpenMVS Contributors. OpenMVS. https://github.com/cdcseacave/openMVS, 2026. Accessed: 2026-05-06.
- [37]L. Pan, D. Baráth, M. Pollefeys, and J. L. Schönberger. Global Structure-From-Motion Revisited. In European Conference on Computer Vision, pages 58–77, 2024.arxiv.org/abs/2407.20219
- [38]Pix4D. Pix4D, 2012. https://www.pix4d.com/.
- [39]Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Proceedings of the 38th International Conference on Machine Learning, pages 8748–8763. PMLR, 2021.arxiv.org/abs/2103.00020
- [40]Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3D: Large-scale learning and evaluation of real-life 3D category reconstruction. In International Conference on Computer Vision, pages 10881–10891, 2021.
- [41]Weining Ren, Xiao Tan, and Kai Han. Speed3r: Sparse feed-forward 3d reconstruction models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Findings, 2026.arxiv.org/abs/2603.08055
- [42]Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M. Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In International Conference on Computer Vision, 2021.arxiv.org/abs/2011.02523
- [43]Aurko Roy, Mohammad Saffar, Ashish Vaswani, and David Grangier. Efficient content-based sparse attention with routing transformers. Transactions of the Association for Computational Linguistics, 9:53–68, 2021.
- [44]P.E. Sarlin, D. DeTone, T. Malisiewicz, and A. Rabinovich. SuperGlue: Learning Feature Matching with Graph Neural Networks. In Conference on Computer Vision and Pattern Recognition, 2020.arxiv.org/abs/1911.11763
- [45]P.-E. Sarlin, M. Dusmanu, J. L. Schönberger, P. Speciale, L. Gruber, V. Larsson, O. Miksik, and M. Pollefeys. LaMAR: Benchmarking Localization and Mapping for Augmented Reality. In European Conference on Computer Vision, 2022.arxiv.org/abs/2210.10770
- [46]J.L. Schönberger and J.M. Frahm. Structure-From-Motion Revisited. In Conference on Computer Vision and Pattern Recognition, 2016.DOI
- [47]T. Schöps, J. Schönberger, S. Gallian, T. Sattler, K. Schindler, M. Pollefeys, and A. Geiger. A Multi-View Stereo Benchmark with High-Resolution Images and Multi-Camera Videos. In Conference on Computer Vision and Pattern Recognition, 2017.DOI
- [48]You Shen, Zhipeng Zhang, Yansong Qu, Xiawu Zheng, Jiayi Ji, Shengchuan Zhang, and Lijuan Cao. FastVGGT: Fast visual geometry transformer. In International Conference on Learning Representations, 2026.
- [49]J. Shotton, B. Glocker, C. Zach, S. Izadi, A. Criminisi, and A. Fitzgibbon. Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images. In Conference on Computer Vision and Pattern Recognition, pages 2930–2937, 2013.DOI
- [50]Julian Straub, Thomas Whelan, Lingni Ma, Yufan Chen, Erik Wijmans, Simon Green, Jakob J. Engel, Raul Mur-Artal, Carl Ren, Shobhit Verma, et al. The Replica dataset: A digital replica of indoor spaces. arXiv preprint arXiv:1906.05797, 2019.arxiv.org/abs/1906.05797
- [51]J. Sun, Z. Shen, Y. Wang, H. Bao, and X. Zhou. LoFTR: Detector-Free Local Feature Matching with Transformers. In Conference on Computer Vision and Pattern Recognition, 2021.arxiv.org/abs/2104.00680
- [52]Xianbing Sun, Zhikai Zhu, Zhengyu Lou, Bo Yang, Jinyang Tang, Liqing Zhang, He Wang, and Jianfu Zhang. Avggt: Rethinking global attention for accelerating vggt. In Conference on Computer Vision and Pattern Recognition, pages 251–260, 2026.arxiv.org/abs/2512.02541
- [53]E. Tola, C. Strecha, and P. Fua. Efficient Large Scale Multi-View Stereo for Ultra High Resolution Image Sets. Machine Vision and Applications, 23(5), 2012.DOI
- [54]Fabio Tosi, Yiyi Liao, Carolin Schmitt, and Andreas Geiger. SMD-Nets: Stereo mixture density networks. In Conference on Computer Vision and Pattern Recognition, pages 8938–8948, 2021.arxiv.org/abs/2104.03866
- [55]Michał Tyszkiewicz, Pascal Fua, and Eduard Trulls. Disk: Learning local features with policy gradient. Advances in neural information processing systems, 33:14254–14265, 2020.arxiv.org/abs/2006.13566
- [56]S. Umeyama. Least-Squares Estimation of Transformation Parameters Between Two Point Patterns. IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(4), 1991.DOI
- [57]Basile Van Hoorick, Rundi Wu, Ege Ozguroglu, Kyle Sargent, Ruoshi Liu, Pavel Tokmakov, Achal Dave, Changxi Zheng, and Carl Vondrick. Generative camera dolly: Extreme monocular dynamic novel view synthesis. In European Conference on Computer Vision, pages 313–331, 2024.arxiv.org/abs/2405.14868
- [58]Chung-Shien Brian Wang, Christian Schmidt, Jens Piekenbrinck, and Bastian Leibe. Block-sparse global attention for efficient multi-view geometry transformers. In Conference on Computer Vision and Pattern Recognition, pages 14546–14555, 2026.arxiv.org/abs/2509.07120
- [59]Hengyi Wang and Lourdes Agapito. 3D reconstruction with spatial memory. In International Conference on 3D Vision, 2025.arxiv.org/abs/2408.16061
- [60]Jianyuan Wang, Minghao Chen, Nikita Karaev, Andrea Vedaldi, Christian Rupprecht, and David Novotny. Vggt: Visual geometry grounded transformer. In Conference on Computer Vision and Pattern Recognition, 2025.arxiv.org/abs/2503.11651
- [61]Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, and Christian Rupprecht. VGGT-Ω. In Conference on Computer Vision and Pattern Recognition, 2026.DOI
- [62]Qianqian Wang, Yifei Zhang, Aleksander Holynski, Alexei A. Efros, and Angjoo Kanazawa. Continuous 3D perception model with persistent state. In Conference on Computer Vision and Pattern Recognition, 2025.arxiv.org/abs/2501.12387
- [63]Shuzhe Wang, Vincent Leroy, Yohann Cabon, Boris Chidlovskii, and Jerome Revaud. Dust3r: Geometric 3d vision made easy. arXiv Preprint, 2312.14132, 2024.arxiv.org/abs/2312.14132
- [64]Wenshan Wang, Delong Zhang, Xiangwei Wang, Yaoyu Hu, Yuheng Qiu, Chen Wang, Yafei Hu, Ashish Kapoor, and Sebastian Scherer. TartanAir: A dataset to push the limits of visual SLAM. In International Conference on Intelligent Robots and Systems, 2020.arxiv.org/abs/2003.14338
- [65]Yifan Wang, Jianjun Zhou, Haoyi Zhu, et al. π³: Scalable permutation-equivariant visual geometry learning. arXiv Preprint, 2507.13347, 2025.
- [66]Zipeng Wang and Dan Xu. Flashvggt: Efficient and scalable visual geometry transformers with compressed descriptor attention. In Conference on Computer Vision and Pattern Recognition, pages 21618–21835, 2026.arxiv.org/abs/2512.01540
- [67]K.Q. Weinberger and L.K. Saul. Fast Solvers and Efficient Implementations for Distance Metric Learning. In International Conference on Machine Learning, 2008.DOI
- [68]Yuhuai Wu, Markus N. Rabe, DeLesley Hutchins, and Christian Szegedy. Memorizing transformers. In International Conference on Learning Representations, 2022.arxiv.org/abs/2203.08913
- [69]Jianfeng Xiang, Zelong Lv, Sicheng Xu, Yu Deng, Ruicheng Wang, Bowen Zhang, Dong Chen, Xin Tong, and Jiaolong Yang. Structured 3d latents for scalable and versatile 3d generation. In Conference on Computer Vision and Pattern Recognition, 2025.arxiv.org/abs/2412.01506
- [70]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. In Conference on Computer Vision and Pattern Recognition, 2025.arxiv.org/abs/2501.13928
- [71]Yao Yao, Zixin Luo, Shiwei Li, Jingyang Zhang, Yufan Ren, Lei Zhou, Tian Fang, and Long Quan. BlendedMVS: A Large-Scale Dataset for Generalized Multi-View Stereo Networks. In Conference on Computer Vision and Pattern Recognition, 2020.arxiv.org/abs/1911.10127
- [72]Manzil Zaheer, Guru Guruganesh, Kumar Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, and Amr Ahmed. Big bird: Transformers for longer sequences. In Advances in Neural Information Processing Systems, 2020.arxiv.org/abs/2007.14062
- [73]Amir R. Zamir, Alexander Sax, William Shen, Leonidas J. Guibas, Jitendra Malik, and Silvio Savarese. Taskonomy: Disentangling task transfer learning. In Conference on Computer Vision and Pattern Recognition, pages 3712–3722, 2018.arxiv.org/abs/1804.08328
- [74]Haoyu Zhang, Zeyu Zhang, Zedong Zhou, Yang Zhao, and Hao Tang. Lite3r: A model-agnostic framework for efficient feed-forward 3d reconstruction. arXiv Preprint, 2605.11354, 2026.arxiv.org/abs/2605.11354
- [75]Dong Zhuo, Wenzhao Zheng, Jiahe Guo, Yuqi Wu, Jie Zhou, and Jiwen Lu. Streaming 4d visual geometry transformer. arXiv Preprint, 2507.11539, 2025.arxiv.org/abs/2507.11539


