Zebra3D: hybrid sparse-linear attention speeds 3D foundation models up to 7.1x
zhenjun_zhao · x · 2026-10-13
The arXiv paper Zebra3D introduces a layerwise sparse-linear hybrid attention framework that makes pretrained feed-forward 3D foundation models efficient without training from scratch.
Approach:
- Across VGGT, HunyuanWorld-Mirror and π³, early layers tolerate linear replacement while sensitive middle-to-late layers show concentrated attention
- Zebra3D assigns linear-centered attention to tolerant stages and tile-sparse attention to sensitive ones, using landmark summaries to keep unselected context
- Closed-form affine fitting enables training-free conversion; optional post-training adds task supervision plus feature distillation
Results:
- Up to 7.1× end-to-end speedup at 1,024 views (π³), 6.2× on VGGT, 5.5× on HYWM
- Zebra3D-VGGT hits 80.5% RealEstate10K AUC@30 at 94% attention sparsity, within 1.1 points of dense
- Best Chamfer distance on ScanNet-50 long-sequence reconstruction (1,000 frames) at 5.6× speedup
More from Research
- Allen AI's TutorMoments Benchmark Finds LLM Tutors Over-Help Students — allen_ai · 2026-10-13
- Ben Goertzel revisits Omohundro's Basic AI Drives, finding six assumptions that decide which minds apply — bengoertzel · 2026-10-13
- Sim-to-real skepticism: pixel-space native training beats generated simulators, argues DeWitte — chris_j_paxton · 2026-10-13
- DeepMind publishes survey mapping theories of machine consciousness — Dr_Atoosa · 2026-10-13
- Noise: a speedrun for pretraining transformers without backprop — madhavsinghal_ · 2026-10-13
- Harvard releases freeinference_agentic_trace, a million-scale agent trace dataset — harvardMadsys · 2026-10-13