Research map / 3D vision / Depth and stereo
VGGT-Ω
Jianyuan Wang, Minghao Chen, Shangzhan Zhang, Nikita Karaev, Johannes Schönberger, Patrick Labatut, Piotr Bojanowski, David Novotny, Andrea Vedaldi, Christian Rupprecht
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 changed | Where | Measured | Effect |
|---|---|---|---|
| scale training data from 2K to 2M sequences (ablation) | VGGT-Omega 1B · point error (1B ablation) | point error: 0.275 -> 0.073 | Improved |
| scale model from 1B to 10B (ablation) | VGGT-Omega · Sintel camera pose | AUC@3: 35.5 -> 40.0 | Improved |
| replace 25% of global attention layers with register attention (ablation) | VGGT-Omega 1B · point error (1B ablation) | point error: 0.071 -> 0.073 | No clear effect |
| replace 25% of global attention with register attention (speed) (ablation) | VGGT-Omega · inference speed | speedup: 20-25% speedup | Improved |
| replace all global attention with register attention (ablation) | VGGT-Omega · 1000-frame inference | runtime seconds: 240.2 -> 11.7 | Improved |
| add point and matching losses to multi-task training (ablation) | VGGT-Omega 1B · point error (1B ablation) | point error: 0.078 -> 0.073 | Improved |
| use VGGT's multi-head multi-task setup instead of single dense head (ablation) | VGGT-Omega 1B · point error (1B ablation) | point error: 0.073 -> 0.070 | No clear effect |
| add 10% self-supervised training steps (ablation) | VGGT-Omega 1B · point error (1B ablation) | point error: 0.073 -> 0.070 | Improved |
| use own annotation pipeline instead of MegaSaM pseudo labels | pseudo ground truth · Sintel annotation quality | AUC@30 camera: 62.1 -> 96.4 | Improved |
| add frozen VGGT-Omega scene tokens to OpenVLA-OFT | OpenVLA-OFT · LIBERO | average success rate: 97.1 -> 98.5 | Improved |
| use VGGT-Omega 10B over DA3 | camera pose · Sintel | AUC@3: 16.2 -> 40.0 | Improved |
| use text-only LLM embedding instead of VLM embedding for retrieval | language-aligned registers · language retrieval | top-1 accuracy: 76.8 -> 47.5 | Hurt |
Closest papers by meaning
- VGGT: Visual Geometry Grounded Transformer (CVPR 2025 · similarity 0.91)
- TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention (NeurIPS 2026 · similarity 0.88)
- IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer (NeurIPS 2026 · similarity 0.88)
- MegaSynth: Scaling Up 3D Scene Reconstruction with Synthesized Data (CVPR 2025 · similarity 0.88)
- Rethinking the State Update Gate for Long-Sequence Recurrent 3D Reconstruction (NeurIPS 2026 · similarity 0.87)
- LiteVGGT: Boosting Vanilla VGGT via Geometry-aware Cached Token Merging (CVPR 2026 · similarity 0.87)
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.
