3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering

Hugging Face Daily Papers Papers

Summary

3DZip is a three-stage token compression framework for 3D vision-language models that uses voxelization, diversity-guided anchor selection, and spatial constraint merging to reduce tokens while preserving spatial reasoning. It retains 94.7% of original performance with only 128 tokens and achieves 1.92x faster inference on 3DQA benchmarks.

Recent 3D vision-language models (3D VLMs) construct geometry aware tokens by projecting 2D visual features into world coordinates, enabling spatial reasoning for tasks such as 3D question answering. However, this design generates thousands of tokens per scene, resulting in substantial computational and memory overhead. While token compression has been extensively studied in 2D VLMs, existing approaches rely on semantic relevance or attention-based selection that overlook the structured spatial nature of 3D tokens. Moreover, redundancy in 3D representations cannot be resolved by spatial proximity alone, as object-level token imbalance persists even after spatial aggregation. To address this, we propose 3DZip, a three-stage token compression framework that first applies coarse voxelization to remove point-level redundancy, then selects anchor tokens based on feature-space diversity via a Determinantal Point Process, and finally merges remaining tokens under spatial constraints to preserve geometric coherence. Experiments on three 3D question answering benchmarks demonstrate that 3DZip consistently outperforms existing compression methods, retaining 94.7% of the original performance with only 128 tokens, achieving a 1.92times faster inference speed.
Original Article
View Cached Full Text

Cached at: 08/04/26, 09:38 AM

Paper page - 3DZip: Spatial-Aware Feature Diversity-Guided Token Compression for 3D Question Answering

Source: https://huggingface.co/papers/2608.01185

Abstract

Recent3Dvision-languagemodels(3DVLMs)constructgeometryawaretokensbyprojecting2Dvisualfeaturesintoworldcoordinates,enablingspatialreasoningfortaskssuchas3Dquestionanswering.However,thisdesigngeneratesthousandsoftokensperscene,resultinginsubstantialcomputationalandmemoryoverhead.Whiletokencompressionhasbeenextensivelystudiedin2DVLMs,existingapproachesrelyonsemanticrelevanceorattention-basedselectionthatoverlookthestructuredspatialnatureof3Dtokens.Moreover,redundancyin3Drepresentationscannotberesolvedbyspatialproximityalone,asobject-leveltokenimbalancepersistsevenafterspatialaggregation.Toaddressthis,wepropose3DZip,athree-stagetokencompressionframeworkthatfirstappliescoarsevoxelizationtoremovepoint-levelredundancy,thenselectsanchortokensbasedonfeature-spacediversityviaaDeterminantalPointProcess,andfinallymergesremainingtokensunderspatialconstraintstopreservegeometriccoherence.Experimentsonthree3Dquestionansweringbenchmarksdemonstratethat3DZipconsistentlyoutperformsexistingcompressionmethods,retaining94.7%oftheoriginalperformancewithonly128tokens,achievinga1.92timesfasterinferencespeed.

View arXiv pageView PDFProject pageGitHub13Add to collection

Get this paper in your agent:

hf papers read 2608\.01185

Don’t have the latest CLI?curl \-LsSf https://hf\.co/cli/install\.sh \| bash

Models citing this paper0

No model linking this paper

Cite arxiv.org/abs/2608.01185 in a model README.md to link it from this page.

Datasets citing this paper0

No dataset linking this paper

Cite arxiv.org/abs/2608.01185 in a dataset README.md to link it from this page.

Spaces citing this paper0

No Space linking this paper

Cite arxiv.org/abs/2608.01185 in a Space README.md to link it from this page.

Collections including this paper0

No Collection including this paper

Add this paper to acollectionto link it from this page.

Similar Articles

ZipTok3D: High-Fidelity 3D Tokenization with Compact Token Prefixes

Hugging Face Daily Papers

ZipTok3D is a 3D tokenizer that achieves high-fidelity reconstruction from extremely short token sequences by organizing geometry into compact global-token prefixes and using iterative decoding, outperforming baselines with significantly reduced token lengths.