Research map / 3D vision / 3D generation

TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention

NeurIPS 2026 · 3D vision › 3D generation · approach: Learning 3D reconstruction from images · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes TurboVGGT, a feed-forward visual geometry transformer that replaces VGGT's dense global attention with adaptive sparse global attention (frame-wise gating over three sparsity branches plus learned compressed representative tokens), achieving multi-fold speedups with comparable reconstruction quality.

Problem

Visual geometry transformers for multi-view 3D reconstruction (e.g., VGGT) have quadratic cost in global attention over tokens across frames, limiting speed and scalability; existing efficient variants use fixed sparsity across layers and frames and lack learned representative tokens.

Approach

Route each frame via a gating network to one of three sparsity branches (75%, 89%, 94%) per layer, and in each branch learn a weight matrix that compresses each frame's patch tokens into a small set of representative tokens; global cross-attention then runs between these compressed tokens and all dense tokens, alternated with frame attention.

Key results

• On 7-Scenes, TurboVGGT reaches 9.6s inference vs. VGGT's 38.1s with AUC@30 81.87 vs. 77.76 and AbsRel 0.296 vs. 0.264 (VGGT better on AbsRel).
• On ScanNet average over 100/300/500 frames, TurboVGGT achieves the lowest average CD of 0.410 and 10.7s average time, versus FastVGGT 0.442 and 14.5s.
• Peak inference memory on 7-Scenes is 23.47 GB for TurboVGGT vs. 25.24 GB for VGGT; claimed 7x/11x/18x speedup over VGGT at 1000 frames for variants.

Surprising findings

• TurboVGGT's AbsRel on 7-Scenes (0.296) is worse than VGGT's (0.264), and its Acc is slightly worse than some baselines on N-RGBD.

Limitations

• Does not explicitly model temporal dynamics, limiting performance on highly dynamic scenes.
• Room for improvement on challenging scenes with complex geometry.

Experimental setup

Datasets: BlendedMVS Mapillary Planet-Scale Depth ScanNet++ v2 Spring TartanAirV2-WB UnrealStereo4K Aria Synthetic Environments DL3DV-10K Dynamic Replica MegaDepth MVS-Synth ParallelDomain-4D

Benchmarks: 7-Scenes N-RGBD ScanNet RealEstate10K Sintel Bonn

Metrics: Accuracy Completeness Normal Consistency Chamfer Distance RRA@30 RTA@30 AUC@30 AbsRel delta<1.25 inference time

Baselines: VGGT Fast3R SparseVGGT AVGGT FastVGGT FLARE Speed3R FlashVGGT π3 MapAnything

Builds on: VGGT MapAnything π3 DINOv2

Closest papers by meaning

• VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.90)
• RegimeVGGT: Layer-Wise Spatially Preserving Redundancy Removal for Visual Geometry Grounded Transformer (NeurIPS 2026 · similarity 0.89)
• LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging (CVPR 2026 · similarity 0.88)
• Good Token Hunting: A Hitchhiker's Guide to Token Selection for Visual Geometry Transformers (NeurIPS 2026 · similarity 0.88)
• Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.87)
• FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry (NeurIPS 2026 · similarity 0.87)