{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:C2JAM3NK5YGL7RX6FLMB73WJOI","short_pith_number":"pith:C2JAM3NK","schema_version":"1.0","canonical_sha256":"1692066daaee0cbfc6fe2ad81feec9723902bb4d9e73a88769cb1904398fc75b","source":{"kind":"arxiv","id":"2604.03118","version":2},"attestation_state":"computed","paper":{"title":"Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Salt distills video models to 2-4 steps by regularizing the endpoint consistency of consecutive denoising updates and conditioning on KV cache states.","cross_cats":["eess.IV"],"primary_cat":"cs.CV","authors_text":"Bingqi Ma, Dailan He, Guanglu Song, Jun Zhang, Xiahong Wang, Xingtong Ge, Yi Zhang, Yu Liu, Yushi Huang","submitted_at":"2026-04-03T15:43:21Z","abstract_excerpt":"Distilling video generation models to extremely low inference budgets (e.g., 2--4 NFEs) is crucial for real-time deployment, yet remains challenging. Trajectory-style consistency distillation often becomes conservative under complex video dynamics, yielding an over-smoothed appearance and weak motion. Distribution matching distillation (DMD) can recover sharp, mode-seeking samples, but its local training signals do not explicitly regularize how denoising updates compose across timesteps, making composed rollouts prone to drift. To overcome this challenge, we propose Self-Consistent Distributio"},"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":true},"canonical_record":{"source":{"id":"2604.03118","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-04-03T15:43:21Z","cross_cats_sorted":["eess.IV"],"title_canon_sha256":"1f81f90fc8f84eee1cad6c816ad45b6b5266f180f735a92289f264f474f1ece6","abstract_canon_sha256":"2e5b42066cdece165e571d2c17fce0e5582fb712271a3342f2469b7e8aa7f87c"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-02T01:17:30.449463Z","signature_b64":"EWWobVA2m3EfDPBiCfg6yxFl/k8siX+kKv2NfWYdGzvbYd1mcVNNgLwDr/VSHMab601q1MF49DNhDHiSEOBRDA==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1692066daaee0cbfc6fe2ad81feec9723902bb4d9e73a88769cb1904398fc75b","last_reissued_at":"2026-07-02T01:17:30.448942Z","signature_status":"signed_v1","first_computed_at":"2026-07-02T01:17:30.448942Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Salt distills video models to 2-4 steps by regularizing the endpoint consistency of consecutive denoising updates and conditioning on KV cache states.","cross_cats":["eess.IV"],"primary_cat":"cs.CV","authors_text":"Bingqi Ma, Dailan He, Guanglu Song, Jun Zhang, Xiahong Wang, Xingtong Ge, Yi Zhang, Yu Liu, Yushi Huang","submitted_at":"2026-04-03T15:43:21Z","abstract_excerpt":"Distilling video generation models to extremely low inference budgets (e.g., 2--4 NFEs) is crucial for real-time deployment, yet remains challenging. Trajectory-style consistency distillation often becomes conservative under complex video dynamics, yielding an over-smoothed appearance and weak motion. Distribution matching distillation (DMD) can recover sharp, mode-seeking samples, but its local training signals do not explicitly regularize how denoising updates compose across timesteps, making composed rollouts prone to drift. To overcome this challenge, we propose Self-Consistent Distributio"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Across extensive experiments on both non-autoregressive backbones (e.g., Wan 2.1) and autoregressive real-time paradigms (e.g., Self Forcing), our method, dubbed Salt, consistently improves low-NFE video generation quality while remaining compatible with diverse KV-cache memory mechanisms.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That explicitly regularizing the endpoint-consistent composition of consecutive denoising updates will prevent drift in composed rollouts, and that the cache-conditioned feature alignment will reliably steer low-quality outputs toward high-quality references without introducing new inconsistencies.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Salt improves low-step video generation quality by adding endpoint-consistent regularization to distribution matching distillation and using cache-conditioned feature alignment for autoregressive models.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Salt distills video models to 2-4 steps by regularizing the endpoint consistency of consecutive denoising updates and conditioning on KV cache states.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"e61415009b9903b5b935814bcc2518e5c3e72ee3209f655ba4950ae499913a52"},"source":{"id":"2604.03118","kind":"arxiv","version":2},"verdict":{"id":"7e06cb90-3026-4b27-bdb4-05cf9ae4fe50","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-13T19:35:26.286285Z","strongest_claim":"Across extensive experiments on both non-autoregressive backbones (e.g., Wan 2.1) and autoregressive real-time paradigms (e.g., Self Forcing), our method, dubbed Salt, consistently improves low-NFE video generation quality while remaining compatible with diverse KV-cache memory mechanisms.","one_line_summary":"Salt improves low-step video generation quality by adding endpoint-consistent regularization to distribution matching distillation and using cache-conditioned feature alignment for autoregressive models.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That explicitly regularizing the endpoint-consistent composition of consecutive denoising updates will prevent drift in composed rollouts, and that the cache-conditioned feature alignment will reliably steer low-quality outputs toward high-quality references without introducing new inconsistencies.","pith_extraction_headline":"Salt distills video models to 2-4 steps by regularizing the endpoint consistency of consecutive denoising updates and conditioning on KV cache states."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2604.03118/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":2,"snapshot_sha256":"8ec0f5c68f0b1507d7aebce363ae127ab1481e81b2fb6d769b7c3be013fb19c0"},"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":"2604.03118","created_at":"2026-07-02T01:17:30.449000+00:00"},{"alias_kind":"arxiv_version","alias_value":"2604.03118v2","created_at":"2026-07-02T01:17:30.449000+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2604.03118","created_at":"2026-07-02T01:17:30.449000+00:00"},{"alias_kind":"pith_short_12","alias_value":"C2JAM3NK5YGL","created_at":"2026-07-02T01:17:30.449000+00:00"},{"alias_kind":"pith_short_16","alias_value":"C2JAM3NK5YGL7RX6","created_at":"2026-07-02T01:17:30.449000+00:00"},{"alias_kind":"pith_short_8","alias_value":"C2JAM3NK","created_at":"2026-07-02T01:17:30.449000+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2605.27336","citing_title":"PARE: Pruning and Adaptive Routing for Efficient Video Generation","ref_index":8,"is_internal_anchor":true},{"citing_arxiv_id":"2605.23458","citing_title":"One-Forcing: Towards Stable One-Step Autoregressive Video Generation","ref_index":41,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":2,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI","json":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI.json","graph_json":"https://pith.science/api/pith-number/C2JAM3NK5YGL7RX6FLMB73WJOI/graph.json","events_json":"https://pith.science/api/pith-number/C2JAM3NK5YGL7RX6FLMB73WJOI/events.json","paper":"https://pith.science/paper/C2JAM3NK"},"agent_actions":{"view_html":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI","download_json":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI.json","view_paper":"https://pith.science/paper/C2JAM3NK","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2604.03118&json=true","fetch_graph":"https://pith.science/api/pith-number/C2JAM3NK5YGL7RX6FLMB73WJOI/graph.json","fetch_events":"https://pith.science/api/pith-number/C2JAM3NK5YGL7RX6FLMB73WJOI/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI/action/timestamp_anchor","attest_storage":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI/action/storage_attestation","attest_author":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI/action/author_attestation","sign_citation":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI/action/citation_signature","submit_replication":"https://pith.science/pith/C2JAM3NK5YGL7RX6FLMB73WJOI/action/replication_record"}},"created_at":"2026-07-02T01:17:30.449000+00:00","updated_at":"2026-07-02T01:17:30.449000+00:00"}