Research map / Architectures and efficiency / Attention and transformers
SAF3R: Dynamic Sparse Attention for Feed-Forward 3D Reconstruction Transformers
Jianing Deng, Yuanzhe LI, Jialu Wang, Song Wang, Tianlong Chen, Huanrui Yang, Jingtong Hu
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 |
|---|---|---|---|
| add offline head profiling with tailored kernels (+Dynamic) over uniform sampling (ablation) | SAF3R on VGGT · 7Scenes | AUC@3: 22.6 -> 24.2 (sparsity 0.85 -> 0.91) | Improved |
| add static kernel for position heads over uniform sampling (ablation) | SAF3R on VGGT · 7Scenes | AUC@3: 22.6 -> 23.3 (sparsity 0.85 -> 0.90) | Improved |
| add offline head profiling (+Dynamic) over uniform sampling (ablation) | SAF3R on DA3 · HiRoom | AUC@3: 72.5 -> 81.8 | Improved |
| add offline head profiling (+Dynamic) over uniform sampling (ablation) | SAF3R on VGGT · HiRoom | AUC@30: 82.9 -> 78.8 | Hurt |
| add offline head profiling (+Dynamic) over uniform sampling (ablation) | SAF3R on VGGT · five datasets (DTU, 7Scenes, ETH3D, ScanNet++) | AUC@3: DTU 71.7 -> 74.7; HiRoom 41.2 -> 39.1 | Mixed |
| use DINO Top-K (K=4) instead of uniform stride s=32 on correspondence heads (ablation) | VGGT layer 14 · Co3D-v2 | AUC@3: 43.77 -> 48.05 | Improved |
| increase DINO Top-K from K=2 to K=4 (ablation) | VGGT layer 14 · 7Scenes | AUC@3: 22.92 -> 23.34 | Improved |
| skip correspondence heads entirely (ablation) | VGGT layer 14 · Co3D-v2 | AUC@30: 90.34 -> 19.55 | Hurt |
| increase uniform stride sampling from 1.6% to 3.1% keys (s=64 to s=32) (ablation) | VGGT layer 14 · 7Scenes | AUC@3: 19.94 -> 22.64 | Improved |
| use DINO Top-K (K=4) instead of uniform stride s=32 (ablation) | DA3 layer 7 · Co3D-v2 | AUC@3: 51.74 -> 52.48 | Improved |
| SAF3R instead of ADA3 sparse attention | DA3-Giant · ScanNet-50 dense 300 images | ARE: 12.43 -> 5.19 | Improved |
| increase K for DINO Top-K on large outdoor scene with loop closures (ablation) | DA3 · Oxford Spires | trajectory accuracy | Improved |
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.
- FastVGGT — FastVGGT: Fast Visual Geometry Transformer (ICLR 2026) · better on 2 of 3 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
- Understanding Multi-View Transformers (NeurIPS 2026 · similarity 0.88)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.88)
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.87)
- Fused View-Time Attention and Feedforward Reconstruction for 4D Scene Generation (NeurIPS 2025 · similarity 0.87)
- TurboVGGT: Fast Visual Geometry Reconstruction with Adaptive Alternating Attention (NeurIPS 2026 · similarity 0.87)
- 4D3R: Motion-Aware Neural Reconstruction and Rendering of Dynamic Scenes from Monocular Videos (NeurIPS 2025 · 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.
