3DZip: Efficient Token Compression Framework for 3D Question Answering
Changwoo Baek · hf · 2026-08-04
Current 3D vision-language models (3D VLMs) construct geometry-aware tokens by projecting 2D visual features into world coordinates, generating thousands of tokens per scene and incurring massive computational and memory overhead.
To address this, researchers propose 3DZip, a three-stage token compression framework:
- Coarse voxelization: Removes point-level redundancy.
- Anchor selection: Selects anchor tokens based on feature-space diversity via a Determinantal Point Process (DPP).
- Spatially constrained merging: Merges remaining tokens under spatial constraints to preserve geometric coherence.
Experiments on three 3D QA benchmarks show that 3DZip retains 94.7% of original performance with only 128 tokens, achieving a 1.92x speedup in inference.
More from Research
- Weekend Project: RL-Trained 4B LLM Rewrites AI Text to Fool Open-Source Detectors — matthen2 · 2026-08-04
- NUS Creates Octopus-Inspired Swimming Robot Driven by Just Two Motors — lukas_m_ziegler · 2026-08-04
- Native Tool Calling and Correct Sampling Params Boost LLM Evals by >30 Points — xeophon · 2026-08-04
- KlingTeam Proposes Cross-Category Video Motion Transfer Framework with New Dataset — KlingTeam · 2026-08-04
- DeepVoyager-VL: Building Long-Horizon Multimodal Agents Without RL — Huanyao Zhang · 2026-08-04
- Peking University Introduces ReBA for Vision-Language MoE Load Balancing — PekingUniversity · 2026-08-04