AI Paper Atlas

Research map / 3D vision / 3D generation

Streaming Visual Geometry Transformer

Dong Zhuo, Wenzhao Zheng, Jiahe Guo, Yuqi Wu, Jie Zhou, Jiwen Lu

ICLR 2026 · 3D vision › 3D generation · approach: Learning 3D reconstruction from images

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add knowledge distillation from VGGT teacher to the causal student (ablation)StreamVGGT · 7-Scenes (sparse views)Acc mean: 0.202 -> 0.129Improved
add knowledge distillation (ablation)StreamVGGT · NRGBD (sparse views)Acc mean: 0.189 -> 0.084Improved
add cached memory tokens at inference (ablation)StreamVGGT · Online inference at frame 5inference time: 850.699 ms -> 88.193 msImproved
add FlashAttention-2 (ablation)StreamVGGT · Online inference at frame 5peak memory: 5.4 GB -> 2.3 GBImproved
use cached tokens instead of full-sequence causal attention (ablation)StreamVGGT · ETH3D (streaming)Overall error / time: time 4.7 s -> 0.07 s; Overall 0.753 -> 0.753Improved
windowed streaming (50 frames) for long sequences (ablation)StreamVGGT · 200-frame sequenceinference time: 733.083 ms -> 219.454 ms; Acc mean 0.058 -> 0.054Improved
K-nearest-frames caching with K=50 instead of K=100 (ablation)StreamVGGT · 200-frame sequenceAcc mean: 0.054 -> 0.157Hurt
use causal streaming model instead of offline VGGTStreamVGGT · 7-Scenes (sparse views)Acc mean: 0.088 -> 0.129Hurt
use StreamVGGT instead of CUT3Rstreaming reconstruction · ETH3D (10 frames)Overall: 1.411 -> 0.577Improved
use StreamVGGT instead of CUT3Rstreaming video depth · KITTI video depthAbs Rel: 0.118 -> 0.173Hurt
input noise via random masking (ablation)StreamVGGT · 7-ScenesAcc mean: 0.034 -> 0.054 (0% -> 25% noise)Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy