{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2024:DYT2I4DZGLZQNHRFIK44DNROBU","short_pith_number":"pith:DYT2I4DZ","schema_version":"1.0","canonical_sha256":"1e27a4707932f3069e2542b9c1b62e0d3666287c158523b98d2fbde53849c912","source":{"kind":"arxiv","id":"2408.09000","version":2},"attestation_state":"computed","paper":{"title":"Classifier-Free Guidance is a Predictor-Corrector","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CV"],"primary_cat":"cs.LG","authors_text":"Arwen Bradley, Preetum Nakkiran","submitted_at":"2024-08-16T20:00:55Z","abstract_excerpt":"We investigate the theoretical foundations of classifier-free guidance (CFG). CFG is the dominant method of conditional sampling for text-to-image diffusion models, yet unlike other aspects of diffusion, it remains on shaky theoretical footing. In this paper, we disprove common misconceptions, by showing that CFG interacts differently with DDPM (Ho et al., 2020) and DDIM (Song et al., 2021), and neither sampler with CFG generates the gamma-powered distribution $p(x|c)^\\gamma p(x)^{1-\\gamma}$. Then, we clarify the behavior of CFG by showing that it is a kind of predictor-corrector method (Song "},"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":"2408.09000","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.LG","submitted_at":"2024-08-16T20:00:55Z","cross_cats_sorted":["cs.AI","cs.CV"],"title_canon_sha256":"1a1d5365df5ab6454466fdc925aa1faaa6bbb5efea4c79779b3eb552158009eb","abstract_canon_sha256":"723070a7837b45e0739b616223a3d190b4787d170961aca13dc8b36bfb874969"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-07-05T08:58:26.059936Z","signature_b64":"Tt8XyR1VgnCLfPIC9ECuJWhptOOvcGy9kENtpIIROzb8Ni3d6wpFchOjhoi1JVruTl9N+AilY7SW4yvgQTEEBg==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"1e27a4707932f3069e2542b9c1b62e0d3666287c158523b98d2fbde53849c912","last_reissued_at":"2026-07-05T08:58:26.059464Z","signature_status":"signed_v1","first_computed_at":"2026-07-05T08:58:26.059464Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"Classifier-Free Guidance is a Predictor-Corrector","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"","cross_cats":["cs.AI","cs.CV"],"primary_cat":"cs.LG","authors_text":"Arwen Bradley, Preetum Nakkiran","submitted_at":"2024-08-16T20:00:55Z","abstract_excerpt":"We investigate the theoretical foundations of classifier-free guidance (CFG). CFG is the dominant method of conditional sampling for text-to-image diffusion models, yet unlike other aspects of diffusion, it remains on shaky theoretical footing. In this paper, we disprove common misconceptions, by showing that CFG interacts differently with DDPM (Ho et al., 2020) and DDIM (Song et al., 2021), and neither sampler with CFG generates the gamma-powered distribution $p(x|c)^\\gamma p(x)^{1-\\gamma}$. Then, we clarify the behavior of CFG by showing that it is a kind of predictor-corrector method (Song "},"claims":{"count":0,"items":[],"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"source":{"id":"2408.09000","kind":"arxiv","version":2},"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/2408.09000/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":"2408.09000","created_at":"2026-07-05T08:58:26.059518+00:00"},{"alias_kind":"arxiv_version","alias_value":"2408.09000v2","created_at":"2026-07-05T08:58:26.059518+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2408.09000","created_at":"2026-07-05T08:58:26.059518+00:00"},{"alias_kind":"pith_short_12","alias_value":"DYT2I4DZGLZQ","created_at":"2026-07-05T08:58:26.059518+00:00"},{"alias_kind":"pith_short_16","alias_value":"DYT2I4DZGLZQNHRF","created_at":"2026-07-05T08:58:26.059518+00:00"},{"alias_kind":"pith_short_8","alias_value":"DYT2I4DZ","created_at":"2026-07-05T08:58:26.059518+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":7,"internal_anchor_count":1,"sample":[{"citing_arxiv_id":"2607.07665","citing_title":"Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance","ref_index":1,"is_internal_anchor":true},{"citing_arxiv_id":"2606.10385","citing_title":"Beyond Absolute Imitation: Anchored Residual Guidance for Privileged On-Policy Distillation","ref_index":2,"is_internal_anchor":false},{"citing_arxiv_id":"2606.06624","citing_title":"Principles and Practice of Deep Representation Learning: or a Mathematical Theory of Memory","ref_index":6,"is_internal_anchor":false},{"citing_arxiv_id":"2504.13109","citing_title":"UniEdit-Flow: Unleashing Inversion and Editing in the Era of Flow Models","ref_index":9,"is_internal_anchor":false},{"citing_arxiv_id":"2509.23727","citing_title":"AudioMoG: Guiding Audio Generation with Mixture-of-Guidance","ref_index":3,"is_internal_anchor":false},{"citing_arxiv_id":"2412.06264","citing_title":"Flow Matching Guide and Code","ref_index":10,"is_internal_anchor":false},{"citing_arxiv_id":"2604.08828","citing_title":"Post-Hoc Guidance for Consistency Models by Joint Flow Distribution Learning","ref_index":2,"is_internal_anchor":false}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU","json":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU.json","graph_json":"https://pith.science/api/pith-number/DYT2I4DZGLZQNHRFIK44DNROBU/graph.json","events_json":"https://pith.science/api/pith-number/DYT2I4DZGLZQNHRFIK44DNROBU/events.json","paper":"https://pith.science/paper/DYT2I4DZ"},"agent_actions":{"view_html":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU","download_json":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU.json","view_paper":"https://pith.science/paper/DYT2I4DZ","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2408.09000&json=true","fetch_graph":"https://pith.science/api/pith-number/DYT2I4DZGLZQNHRFIK44DNROBU/graph.json","fetch_events":"https://pith.science/api/pith-number/DYT2I4DZGLZQNHRFIK44DNROBU/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU/action/timestamp_anchor","attest_storage":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU/action/storage_attestation","attest_author":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU/action/author_attestation","sign_citation":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU/action/citation_signature","submit_replication":"https://pith.science/pith/DYT2I4DZGLZQNHRFIK44DNROBU/action/replication_record"}},"created_at":"2026-07-05T08:58:26.059518+00:00","updated_at":"2026-07-05T08:58:26.059518+00:00"}