{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2025:X2PB6AGQZXVG2ASW34KDZ5GHAC","short_pith_number":"pith:X2PB6AGQ","schema_version":"1.0","canonical_sha256":"be9e1f00d0cdea6d0256df143cf4c700897c3c86ad8d21b8fce4a85bee5376b1","source":{"kind":"arxiv","id":"2505.16792","version":1},"attestation_state":"computed","paper":{"title":"REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Kaipeng Zhang, Kai Wang, Mingjia Shi, Pengfei Zhou, Wangbo Zhao, Xuanlei Zhao, Yang You, Yuhao Zhou, Zekai Li, Zhangyang Wang, Zhiyuan Liang, Ziqiao Wang","submitted_at":"2025-05-22T15:34:33Z","abstract_excerpt":"Diffusion Transformers (DiTs) deliver state-of-the-art image quality, yet their training remains notoriously slow. A recent remedy -- representation alignment (REPA) that matches DiT hidden features to those of a non-generative teacher (e.g. DINO) -- dramatically accelerates the early epochs but plateaus or even degrades performance later. We trace this failure to a capacity mismatch: once the generative student begins modelling the joint data distribution, the teacher's lower-dimensional embeddings and attention patterns become a straitjacket rather than a guide. We then introduce HASTE (Holi"},"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":false,"formal_links_present":false},"canonical_record":{"source":{"id":"2505.16792","kind":"arxiv","version":1},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2025-05-22T15:34:33Z","cross_cats_sorted":["cs.AI"],"title_canon_sha256":"1ad3cd38bfc1e18e199927c6e471d8a7962069dd2066b0ddc2e2c6a5d507b0df","abstract_canon_sha256":"226321d23a536880ff9c1af94795f322ebf8ad6a632ed8c0eed4b39055c315eb"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T11:07:41.238696Z","signature_b64":"rmNYAp8WMO6H8MauLCIEVflreKOHaxiiC4WVIBA9UcHdbDgQVYOQqeNi2PVmRMBVYt/RMflbTO8Ypi5FOYpbBw==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"be9e1f00d0cdea6d0256df143cf4c700897c3c86ad8d21b8fce4a85bee5376b1","last_reissued_at":"2026-07-05T11:07:41.238147Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T11:07:41.238147Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"REPA Works Until It Doesn't: Early-Stopped, Holistic Alignment Supercharges Diffusion Training","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Kaipeng Zhang, Kai Wang, Mingjia Shi, Pengfei Zhou, Wangbo Zhao, Xuanlei Zhao, Yang You, Yuhao Zhou, Zekai Li, Zhangyang Wang, Zhiyuan Liang, Ziqiao Wang","submitted_at":"2025-05-22T15:34:33Z","abstract_excerpt":"Diffusion Transformers (DiTs) deliver state-of-the-art image quality, yet their training remains notoriously slow. A recent remedy -- representation alignment (REPA) that matches DiT hidden features to those of a non-generative teacher (e.g. DINO) -- dramatically accelerates the early epochs but plateaus or even degrades performance later. We trace this failure to a capacity mismatch: once the generative student begins modelling the joint data distribution, the teacher's lower-dimensional embeddings and attention patterns become a straitjacket rather than a guide. We then introduce HASTE (Holi"},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2505.16792","kind":"arxiv","version":1},"verdict":{"id":null,"model_set":{},"created_at":null,"strongest_claim":"","one_line_summary":"","pipeline_version":null,"weakest_assumption":"","pith_extraction_headline":""},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2505.16792/integrity.json","findings":[],"available":true,"detectors_run":[],"snapshot_sha256":"c28c3603d3b5d939e8dc4c7e95fa8dfce3d595e45f758748cecf8e644a296938"},"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"},"aliases":[{"alias_kind":"arxiv","alias_value":"2505.16792","created_at":"2026-07-05T11:07:41.238207+00:00"},{"alias_kind":"arxiv_version","alias_value":"2505.16792v1","created_at":"2026-07-05T11:07:41.238207+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2505.16792","created_at":"2026-07-05T11:07:41.238207+00:00"},{"alias_kind":"pith_short_12","alias_value":"X2PB6AGQZXVG","created_at":"2026-07-05T11:07:41.238207+00:00"},{"alias_kind":"pith_short_16","alias_value":"X2PB6AGQZXVG2ASW","created_at":"2026-07-05T11:07:41.238207+00:00"},{"alias_kind":"pith_short_8","alias_value":"X2PB6AGQ","created_at":"2026-07-05T11:07:41.238207+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":4,"internal_anchor_count":0,"sample":[{"citing_arxiv_id":"2607.02508","citing_title":"From SRA to Self-Flow: Data Augmentation or Self-Supervision?","ref_index":37,"is_internal_anchor":false},{"citing_arxiv_id":"2606.08788","citing_title":"MaskAlign: Token-Subset Representation Alignment for Efficient Diffusion Training","ref_index":26,"is_internal_anchor":false},{"citing_arxiv_id":"2606.00514","citing_title":"Generate in Reconstruction Space, Match in Semantic Space: Transport Geometry for One-Step Generation","ref_index":44,"is_internal_anchor":false},{"citing_arxiv_id":"2604.04050","citing_title":"TORA: Topological Representation Alignment for 3D Shape Assembly","ref_index":50,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC","json":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC.json","graph_json":"https://pith.science/api/pith-number/X2PB6AGQZXVG2ASW34KDZ5GHAC/graph.json","events_json":"https://pith.science/api/pith-number/X2PB6AGQZXVG2ASW34KDZ5GHAC/events.json","paper":"https://pith.science/paper/X2PB6AGQ"},"agent_actions":{"view_html":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC","download_json":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC.json","view_paper":"https://pith.science/paper/X2PB6AGQ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2505.16792&json=true","fetch_graph":"https://pith.science/api/pith-number/X2PB6AGQZXVG2ASW34KDZ5GHAC/graph.json","fetch_events":"https://pith.science/api/pith-number/X2PB6AGQZXVG2ASW34KDZ5GHAC/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC/action/timestamp_anchor","attest_storage":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC/action/storage_attestation","attest_author":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC/action/author_attestation","sign_citation":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC/action/citation_signature","submit_replication":"https://pith.science/pith/X2PB6AGQZXVG2ASW34KDZ5GHAC/action/replication_record"}},"created_at":"2026-07-05T11:07:41.238207+00:00","updated_at":"2026-07-05T11:07:41.238207+00:00"}