{"paper":{"title":"Spark3R: Asymmetric Token Reduction Makes Fast Feed-Forward 3D Reconstruction","license":"http://creativecommons.org/licenses/by-sa/4.0/","headline":"Asymmetric compression of query versus key-value tokens speeds feed-forward 3D reconstruction up to 28 times without retraining.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Haijie Li, Jian Zhang, Jiaqi Zhang, Jiaye Fu, Qiankun Gao, Siwei Ma, Yanmin Wu, Zecheng Tang","submitted_at":"2026-05-07T13:45:37Z","abstract_excerpt":"Feed-forward 3D reconstruction models based on Vision Transformers can directly estimate scene geometry and camera poses from a small set of input images, but scaling them to video inputs with hundreds or thousands of frames remains challenging due to the quadratic cost of global attention layers. Recent token-merging methods accelerate these models by compressing the token sequence within the global attention layers, but they apply a uniform reduction to query tokens and key-value tokens, ignoring their functionally distinct roles in 3D reconstruction. In this work, we identify a key property"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Spark3R achieves up to 28× speedup on 1,000-frame inputs while maintaining competitive reconstruction quality by decoupling the compression of query tokens and key-value tokens with intra-group merging and lightweight pruning.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That query tokens encode view-specific geometric requests sensitive to compression while key-value tokens represent shared scene context tolerant to aggressive compression, and that this functional distinction generalizes across pretrained models and input sequences without retraining.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Asymmetric token reduction, with distinct merging for queries and pruning for key-values plus layer-wise adaptation, delivers up to 28x speedup on 1000-frame 3D reconstruction inputs while preserving competitive quality.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Asymmetric compression of query versus key-value tokens speeds feed-forward 3D reconstruction up to 28 times without retraining.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1b52a6f3622165dbca41dd4b57a5e92e09236db79c62c5e425b5f200faa43aa3"},"source":{"id":"2605.06270","kind":"arxiv","version":2},"verdict":{"id":"ab2ef0a5-f0b5-477e-aaa2-f566ddaa820d","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T13:36:27.852912Z","strongest_claim":"Spark3R achieves up to 28× speedup on 1,000-frame inputs while maintaining competitive reconstruction quality by decoupling the compression of query tokens and key-value tokens with intra-group merging and lightweight pruning.","one_line_summary":"Asymmetric token reduction, with distinct merging for queries and pruning for key-values plus layer-wise adaptation, delivers up to 28x speedup on 1000-frame 3D reconstruction inputs while preserving competitive quality.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That query tokens encode view-specific geometric requests sensitive to compression while key-value tokens represent shared scene context tolerant to aggressive compression, and that this functional distinction generalizes across pretrained models and input sequences without retraining.","pith_extraction_headline":"Asymmetric compression of query versus key-value tokens speeds feed-forward 3D reconstruction up to 28 times without retraining."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.06270/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T18:31:19.611942Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T12:50:09.833483Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"69fa156b166f40e5eba7fa76e121be7e56b08211501b85eeb40f5a39ecdc900b"},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}