{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:BOVYFUXRPEGMZCOI22DN4CQVBB","short_pith_number":"pith:BOVYFUXR","schema_version":"1.0","canonical_sha256":"0bab82d2f1790ccc89c8d686de0a15087affee2a7a4c89d224d2f297efb2f2c6","source":{"kind":"arxiv","id":"2605.14191","version":1},"attestation_state":"computed","paper":{"title":"CoReDiT: Spatial Coherence-Guided Token Pruning and Reconstruction for Efficient Diffusion Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"CoReDiT prunes redundant tokens in diffusion transformers via spatial coherence scores and reconstructs their outputs from neighbors to cut computation while preserving quality.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Fatih Porikli, Hong Cai, Hsin-Pai Cheng, Shizhong Han, Zhuojin Li","submitted_at":"2026-05-13T23:13:29Z","abstract_excerpt":"Diffusion Transformers (DiTs) deliver remarkable image and video generation quality but incur high computational cost, limiting scalability and on-device deployment. We introduce CoReDiT, a structured token pruning framework for DiTs across vision tasks. CoReDiT uses a linear-time spatial coherence score to estimate local redundancy in the latent token lattice and skips high coherence (redundant) tokens in self-attention. To maintain a dense representation and avoid visual discontinuities, we reconstruct skipped attention outputs via coherence-guided aggregation of spatially neighboring retain"},"verification_status":{"content_addressed":true,"pith_receipt":true,"author_attested":false,"weak_author_claims":0,"strong_author_claims":0,"externally_anchored":false,"storage_verified":false,"citation_signatures":0,"replication_records":0,"graph_snapshot":true,"references_resolved":true,"formal_links_present":false},"canonical_record":{"source":{"id":"2605.14191","kind":"arxiv","version":1},"metadata":{"license":"http://creativecommons.org/licenses/by/4.0/","primary_cat":"cs.CV","submitted_at":"2026-05-13T23:13:29Z","cross_cats_sorted":[],"title_canon_sha256":"5f3f341b8d95770267ac541fa599d623099a1bfc01bbbc1bed784f96d6c05e82","abstract_canon_sha256":"d636a1122738d4a1e98a66b91ae9bd1f4347dc86445d0a5f89793b45c77691d3"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-17T23:39:11.140718Z","signature_b64":"Q2CtQ8w1twG4N6SeaCpXuTncbEXx+VKgBhVCbba3sC8Pg+wLal8izzXpW+tEwE0iBvzD4K5rKd+ZISAqvbe5CQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"0bab82d2f1790ccc89c8d686de0a15087affee2a7a4c89d224d2f297efb2f2c6","last_reissued_at":"2026-05-17T23:39:11.140095Z","signature_status":"signed_v1","first_computed_at":"2026-05-17T23:39:11.140095Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"CoReDiT: Spatial Coherence-Guided Token Pruning and Reconstruction for Efficient Diffusion Transformers","license":"http://creativecommons.org/licenses/by/4.0/","headline":"CoReDiT prunes redundant tokens in diffusion transformers via spatial coherence scores and reconstructs their outputs from neighbors to cut computation while preserving quality.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Fatih Porikli, Hong Cai, Hsin-Pai Cheng, Shizhong Han, Zhuojin Li","submitted_at":"2026-05-13T23:13:29Z","abstract_excerpt":"Diffusion Transformers (DiTs) deliver remarkable image and video generation quality but incur high computational cost, limiting scalability and on-device deployment. We introduce CoReDiT, a structured token pruning framework for DiTs across vision tasks. CoReDiT uses a linear-time spatial coherence score to estimate local redundancy in the latent token lattice and skips high coherence (redundant) tokens in self-attention. To maintain a dense representation and avoid visual discontinuities, we reconstruct skipped attention outputs via coherence-guided aggregation of spatially neighboring retain"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across state-of-the-art diffusion backbones including PixArt-α and MagicDrive-V2, CoReDiT achieves up to 55% self-attention FLOPs reduction and inference speedups of 1.33x on cloud GPUs and 1.72x on mobile NPUs, while maintaining high visual quality.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the linear-time spatial coherence score reliably identifies redundant tokens whose removal and subsequent neighbor-based reconstruction will not introduce perceptible visual artifacts or degrade generation quality across diverse prompts and resolutions.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"CoReDiT reduces self-attention FLOPs in DiTs by up to 55% via linear-time spatial coherence pruning and neighbor-based reconstruction, delivering 1.33x-1.72x speedups with maintained quality.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"CoReDiT prunes redundant tokens in diffusion transformers via spatial coherence scores and reconstructs their outputs from neighbors to cut computation while preserving quality.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"d02da5d9d3feeecb90cbfcce96a112fab47ba5eef61efca4e5fce3b142e4e26f"},"source":{"id":"2605.14191","kind":"arxiv","version":1},"verdict":{"id":"75a9ac45-b170-4228-a88b-2acb23233bed","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-15T04:44:28.788273Z","strongest_claim":"Across state-of-the-art diffusion backbones including PixArt-α and MagicDrive-V2, CoReDiT achieves up to 55% self-attention FLOPs reduction and inference speedups of 1.33x on cloud GPUs and 1.72x on mobile NPUs, while maintaining high visual quality.","one_line_summary":"CoReDiT reduces self-attention FLOPs in DiTs by up to 55% via linear-time spatial coherence pruning and neighbor-based reconstruction, delivering 1.33x-1.72x speedups with maintained quality.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the linear-time spatial coherence score reliably identifies redundant tokens whose removal and subsequent neighbor-based reconstruction will not introduce perceptible visual artifacts or degrade generation quality across diverse prompts and resolutions.","pith_extraction_headline":"CoReDiT prunes redundant tokens in diffusion transformers via spatial coherence scores and reconstructs their outputs from neighbors to cut computation while preserving quality."},"references":{"count":39,"sample":[{"doi":"","year":null,"title":"Token merging for fast sta- ble diffusion","work_id":"c2a6db5b-1e5d-46b2-bd17-02ef709ec2b4","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2022,"title":"Token Merging: Your ViT But Faster","work_id":"528509bc-2611-4e7f-a772-ea14d25b6dae","ref_index":2,"cited_arxiv_id":"2210.09461","is_internal_anchor":true},{"doi":"","year":2020,"title":"nuscenes: A multi- modal dataset for autonomous driving","work_id":"1a9f2e91-86e4-47cd-ad1d-adad75cf9987","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Exploring diffusion transformer designs via grafting","work_id":"53d6cabb-10d6-4ec8-94c8-f4eda9eac668","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Flexdit: Dynamic token density control for diffusion transformer, 2024","work_id":"51d825d6-e2e3-43cb-bfb5-1a09206b7b54","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":39,"snapshot_sha256":"a9bc256b7337589452d0f4bcb96b12a8f1f4a3bf32fbbc106b57656c8cd6c68b","internal_anchors":3},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2605.14191","created_at":"2026-05-17T23:39:11.140207+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.14191v1","created_at":"2026-05-17T23:39:11.140207+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.14191","created_at":"2026-05-17T23:39:11.140207+00:00"},{"alias_kind":"pith_short_12","alias_value":"BOVYFUXRPEGM","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_16","alias_value":"BOVYFUXRPEGMZCOI","created_at":"2026-05-18T12:33:37.589309+00:00"},{"alias_kind":"pith_short_8","alias_value":"BOVYFUXR","created_at":"2026-05-18T12:33:37.589309+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":0,"internal_anchor_count":0,"sample":[]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB","json":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB.json","graph_json":"https://pith.science/api/pith-number/BOVYFUXRPEGMZCOI22DN4CQVBB/graph.json","events_json":"https://pith.science/api/pith-number/BOVYFUXRPEGMZCOI22DN4CQVBB/events.json","paper":"https://pith.science/paper/BOVYFUXR"},"agent_actions":{"view_html":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB","download_json":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB.json","view_paper":"https://pith.science/paper/BOVYFUXR","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.14191&json=true","fetch_graph":"https://pith.science/api/pith-number/BOVYFUXRPEGMZCOI22DN4CQVBB/graph.json","fetch_events":"https://pith.science/api/pith-number/BOVYFUXRPEGMZCOI22DN4CQVBB/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB/action/timestamp_anchor","attest_storage":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB/action/storage_attestation","attest_author":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB/action/author_attestation","sign_citation":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB/action/citation_signature","submit_replication":"https://pith.science/pith/BOVYFUXRPEGMZCOI22DN4CQVBB/action/replication_record"}},"created_at":"2026-05-17T23:39:11.140207+00:00","updated_at":"2026-05-17T23:39:11.140207+00:00"}