Research map / 3D vision / 3D generation
Geometry-Guided 3D Visual Token Pruning for Video-Language Models
Han Li, Zehao Huang, Jiahui Fu, Naiyan Wang, Si Liu
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 inter-voxel spatial diversity pruning (SDP) to VCP (ablation) | Geo3DPruner · 16 frames, 90% pruned | avg performance retained: 89.4% -> 94.0% | Improved |
| add intra-voxel view consistency pruning (VCP) to SDP (ablation) | Geo3DPruner · 16 frames, 90% pruned | avg performance retained: 91.3% -> 94.0% | Improved |
| use iterative SDP instead of random voxel selection (ablation) | Geo3DPruner · 16 frames, 90% pruned | avg performance retained: 85.2% -> 94.0% | Improved |
| use iterative SDP instead of uniform voxel selection (ablation) | Geo3DPruner · 16 frames, 90% pruned | avg performance retained: 85.4% -> 94.0% | Improved |
| use attention from geometry encoder instead of cosine similarity for cross-frame relevance (ablation) | Geo3DPruner · 16 frames, 90% pruned | avg performance retained: 89.9% -> 94.0% | Improved |
| Geo3DPruner instead of VisPruner at 90% pruning | Video-3D LLM · 16 frames | avg performance retained: 84.9% -> 92.1% | Improved |
| Geo3DPruner instead of FastV at 90% pruning | Video-3D LLM · 32 frames | avg performance retained: 82.4% -> 92.0% | Improved |
| Geo3DPruner instead of VisPruner at 90% pruning | Video-3D LLM · 8 frames | avg performance retained: 81.6% -> 88.7% | Improved |
| Geo3DPruner instead of VisPruner on dense captioning at 90% pruning | Video-3D LLM · Scan2Cap C@0.5, 16 frames | CIDEr@0.5: 57.3 -> 80.3 | Improved |
| use 16 frames with 50% pruning instead of 8 frames unpruned at equal 1568 tokens | Geo3DPruner · fixed token budget | avg performance: 100% -> 108% | Improved |
| use 32 frames with 75% pruning instead of 8 frames unpruned at equal 1568 tokens | Geo3DPruner · fixed token budget | avg performance: 100% -> 109% | Improved |
| raise pruning ratio from 60% to 90% | Geo3DPruner · 16 frames | avg performance retained: 98.9% -> 92.1% | Hurt |
Closest papers by meaning
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.91)
- VLM-3R: Vision-Language Models Augmented with Instruction-Aligned 3D Reconstruction (CVPR 2026 · similarity 0.90)
- Beyond 3D VQAs: Injecting 3D Spatial Priors into Vision-Language Models for Enhanced Geometric Reasoning (CVPR 2026 · similarity 0.90)
- Learning from Videos for 3D World: Enhancing MLLMs with 3D Vision Geometry Priors (NeurIPS 2025 · similarity 0.90)
- Merge3D: Efficient 3D Multimodal LLMs via Joint 2D-3D Token Merging (CVPR 2026 · similarity 0.90)
- IVC-Prune: Revealing the Implicit Visual Coordinates in LVLMs for Vision Token Pruning (ICLR 2026 · similarity 0.90)
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.
