Research map / Architectures and efficiency / Attention and transformers
Co-Me: Confidence Guided Token Merging for Visual Geometric Transformers
Yutian Chen, Yuheng Qiu, Ruogu Li, Jay Patrikar, Sebastian Scherer
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 |
|---|---|---|---|
| apply Co-Me token merging to VGGT | VGGT* · RealEstate-10K pose (128 frames) | speedup: 7.97x -> 16.2x | Improved |
| apply Co-Me token merging to VGGT | VGGT* · RealEstate-10K pose (128 frames) | AUC translation @10cm: 0.902 -> 0.869 | Hurt |
| apply Co-Me to VGGT | VGGT · DTU-MVS depth (32 frames) | L1 depth error (cm): 0.89 -> 1.10 | Hurt |
| apply Co-Me to MapAnything | MapAnything · DTU-MVS depth (32 frames) | L1 depth error (cm): 4.59 -> 6.80 | Hurt |
| apply Co-Me to DA3-Giant on single-frame NYUd-v2 | DA3-Giant · NYUd-v2 depth (1 frame) | L1 depth error (m): 0.272 -> 0.463 | Hurt |
| apply Co-Me to pi3 for ETH3D point cloud | pi3 · ETH3D point cloud (16 frames) | completeness (cm): 20.8 -> 18.9 | Improved |
| merge by predicted confidence rather than token similarity (ablation) | Co-Me · DTU-MVS depth (32 frames) | speed-accuracy tradeoff | Improved |
| merge low-confidence tokens instead of pick-one or drop-all (ablation) | Co-Me · DTU-MVS depth (32 frames) | performance degradation: over 10x smaller degradation | Improved |
| add attention bias correction (log n) (ablation) | Co-Me · DTU depth and RE10K pose | L1 error / pose accuracy: 4x lower delta L1 on DTU | Improved |
| replace baseline VGGT attention with FlashAttention (VGGT*) | VGGT · DTU-MVS depth (32 frames) | speedup: 1.00x -> 4.81x | Improved |
| use Co-Me on single-frame input | VGGT · ETH3D (1 frame) | speedup: 1.00x -> 1.13x | No clear effect |
| increase input sequence length | Co-Me + VGGT · speedup vs VGGT | speedup: 4.8x at 32 frames to 21.5x at 512 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- FastVGGT — FastVGGT: Fast Visual Geometry Transformer (ICLR 2026) · better on 4 of 8 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
- 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)
- Merge3D: Efficient 3D Multimodal LLMs via Joint 2D-3D Token Merging (CVPR 2026 · similarity 0.85)
- FrameVGGT: Coherence-Preserving Memory for Bounded Streaming Geometry (NeurIPS 2026 · similarity 0.85)
- Exploring Token Pruning in Vision State Space Models (NeurIPS 2024 · similarity 0.85)
- How Many Tokens Do 3D Point Cloud Transformer Architectures Really Need? (NeurIPS 2025 · similarity 0.85)
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.
