{"paper":{"title":"Improving Classifier-Free Guidance of Flow Matching via Manifold Projection","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Reformulating classifier-free guidance in flow matching as manifold-constrained homotopy optimization reduces sensitivity to guidance scales.","cross_cats":["cs.AI"],"primary_cat":"cs.CV","authors_text":"Chao Wang, Haixia Liu, Jian-Feng Cai, Zhengyi Su","submitted_at":"2026-01-29T15:49:31Z","abstract_excerpt":"Classifier-free guidance (CFG) is a widely used technique for controllable generation in diffusion and flow-based models. Despite its empirical success, CFG relies on a heuristic linear extrapolation that is often sensitive to the guidance scale. In this work, we provide a principled interpretation of CFG through the lens of optimization. We demonstrate that the velocity field in flow matching corresponds to the gradient of a sequence of smoothed distance functions, which guides latent variables toward the scaled target image set. This perspective reveals that the standard CFG formulation is a"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"We reformulate the CFG sampling as a homotopy optimization with a manifold constraint. This formulation necessitates a manifold projection step, which we implement via an incremental gradient descent scheme during sampling.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"That the velocity field in flow matching exactly corresponds to the gradient of a sequence of smoothed distance functions guiding latent variables toward the scaled target image set, making standard CFG merely an approximation whose gap controls sensitivity.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Reformulates CFG sampling in flow matching as homotopy optimization with manifold projection via incremental gradient descent and Anderson acceleration, yielding better fidelity and robustness without retraining.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Reformulating classifier-free guidance in flow matching as manifold-constrained homotopy optimization reduces sensitivity to guidance scales.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"218ada2be6aab54ce68536ae4ac14a9b30af0864b9230229b25bb89e9c662cde"},"source":{"id":"2601.21892","kind":"arxiv","version":2},"verdict":{"id":"2ec256b3-fd97-4606-9ddb-05647fedb311","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-16T09:40:42.595531Z","strongest_claim":"We reformulate the CFG sampling as a homotopy optimization with a manifold constraint. This formulation necessitates a manifold projection step, which we implement via an incremental gradient descent scheme during sampling.","one_line_summary":"Reformulates CFG sampling in flow matching as homotopy optimization with manifold projection via incremental gradient descent and Anderson acceleration, yielding better fidelity and robustness without retraining.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"That the velocity field in flow matching exactly corresponds to the gradient of a sequence of smoothed distance functions guiding latent variables toward the scaled target image set, making standard CFG merely an approximation whose gap controls sensitivity.","pith_extraction_headline":"Reformulating classifier-free guidance in flow matching as manifold-constrained homotopy optimization reduces sensitivity to guidance scales."},"references":{"count":0,"sample":[],"resolved_work":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57","internal_anchors":0},"formal_canon":{"evidence_count":2,"snapshot_sha256":"b7b7a2ff27ae5244c865e08bb4baf6e4acb4c71758fd46b10c733d768075f22b"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}