{"record_type":"pith_number_record","schema_url":"https://pith.science/schemas/pith-number/v1.json","pith_number":"pith:2026:OMZRZRQAXJWPO3XTE6QET63URH","short_pith_number":"pith:OMZRZRQA","schema_version":"1.0","canonical_sha256":"73331cc600ba6cf76ef327a049fb7489ef586457619f73b28b1b16b7efe1a70d","source":{"kind":"arxiv","id":"2605.05204","version":2},"attestation_state":"computed","paper":{"title":"D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Step-distilled diffusion models can learn new concepts through on-policy self-distillation without losing their few-step speed.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Dengyang Jiang, Dongyang Liu, Harry Yang, Mingzhe Zheng, Peng Gao, Qilong Wu, Ruoyi Du, Steven Hoi, Xiangpeng Yang, Xin Jin, Zanyi Wang, Zhen Li","submitted_at":"2026-05-06T17:59:34Z","abstract_excerpt":"The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g, Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for direct continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromise their inherent few-step inference capability. To address this, we propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning. We first find that t"},"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":"2605.05204","kind":"arxiv","version":2},"metadata":{"license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","primary_cat":"cs.CV","submitted_at":"2026-05-06T17:59:34Z","cross_cats_sorted":[],"title_canon_sha256":"228087488510f3779d05703507fcbca4d9c3923489b87ae946c343d8183b7cd9","abstract_canon_sha256":"7aff7bbedc1e8517050125d9d897db7c966ef91866f92ef413ee7747ee0fd44b"},"schema_version":"1.0"},"receipt":{"kind":"pith_receipt","key_id":"pith-v1-2026-05","algorithm":"ed25519","signed_at":"2026-05-20T00:04:34.373609Z","signature_b64":"i1EiSN1cQSm+XPHNxXv63Ji7rP0B8LFVCa8G9uuMgmbSGT5GHiJE7i1nAFJ6fK0LbE8YyG0V+l9eeB2CQU2DAQ==","signed_message":"canonical_sha256_bytes","builder_version":"pith-number-builder-2026-05-17-v1","receipt_version":"0.3","canonical_sha256":"73331cc600ba6cf76ef327a049fb7489ef586457619f73b28b1b16b7efe1a70d","last_reissued_at":"2026-05-20T00:04:34.372866Z","signature_status":"signed_v1","first_computed_at":"2026-05-20T00:04:34.372866Z","public_key_fingerprint":"8d4b5ee74e4693bcd1df2446408b0d54"},"graph_snapshot":{"paper":{"title":"D-OPSD: On-Policy Self-Distillation for Continuously Tuning Step-Distilled Diffusion Models","license":"http://arxiv.org/licenses/nonexclusive-distrib/1.0/","headline":"Step-distilled diffusion models can learn new concepts through on-policy self-distillation without losing their few-step speed.","cross_cats":[],"primary_cat":"cs.CV","authors_text":"Dengyang Jiang, Dongyang Liu, Harry Yang, Mingzhe Zheng, Peng Gao, Qilong Wu, Ruoyi Du, Steven Hoi, Xiangpeng Yang, Xin Jin, Zanyi Wang, Zhen Li","submitted_at":"2026-05-06T17:59:34Z","abstract_excerpt":"The landscape of high-performance image generation models is currently shifting from the inefficient multi-step ones to the efficient few-step counterparts (e.g, Z-Image-Turbo and FLUX.2-klein). However, these models present significant challenges for direct continuous supervised fine-tuning. For example, applying the commonly used fine-tuning technique would compromise their inherent few-step inference capability. To address this, we propose D-OPSD, a novel training paradigm for step-distilled diffusion models that enables on-policy learning during supervised fine-tuning. We first find that t"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"By optimized on the model's own trajectory and under it's own supervision, D-OPSD enables the model to learn new concept, style, etc. without sacrificing the original few-step capacity.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The modern diffusion model where the LLM/VLM serves as the encoder can inherit its encoder's in-context capabilities. This is stated as the key finding that enables treating training as an on-policy self-distillation process.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"D-OPSD enables continuous supervised fine-tuning of few-step diffusion models via on-policy self-distillation where the model acts as both teacher (multimodal context) and student (text-only context) on its own roll-outs.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Step-distilled diffusion models can learn new concepts through on-policy self-distillation without losing their few-step speed.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"0b5f630debbacace2b6c1158d075e455381f0a0c345d6240a8ae11c6e35fc2ee"},"source":{"id":"2605.05204","kind":"arxiv","version":2},"verdict":{"id":"e67d005f-0d90-42e6-866b-839b34035cc1","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-08T17:21:25.679765Z","strongest_claim":"By optimized on the model's own trajectory and under it's own supervision, D-OPSD enables the model to learn new concept, style, etc. without sacrificing the original few-step capacity.","one_line_summary":"D-OPSD enables continuous supervised fine-tuning of few-step diffusion models via on-policy self-distillation where the model acts as both teacher (multimodal context) and student (text-only context) on its own roll-outs.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The modern diffusion model where the LLM/VLM serves as the encoder can inherit its encoder's in-context capabilities. This is stated as the key finding that enables treating training as an on-policy self-distillation process.","pith_extraction_headline":"Step-distilled diffusion models can learn new concepts through on-policy self-distillation without losing their few-step speed."},"integrity":{"clean":true,"summary":{"advisory":0,"critical":0,"by_detector":{},"informational":0},"endpoint":"/pith/2605.05204/integrity.json","findings":[],"available":true,"detectors_run":[{"name":"doi_title_agreement","ran_at":"2026-05-19T21:01:19.701997Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T13:46:29.903248Z","status":"completed","version":"1.0.0","findings_count":0}],"snapshot_sha256":"7a93af0ac77f5e5d70472920ba84c19870df33200af4ca95d6da1186d6da58dd"},"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":"2605.05204","created_at":"2026-05-20T00:04:34.372975+00:00"},{"alias_kind":"arxiv_version","alias_value":"2605.05204v2","created_at":"2026-05-20T00:04:34.372975+00:00"},{"alias_kind":"doi","alias_value":"10.48550/arxiv.2605.05204","created_at":"2026-05-20T00:04:34.372975+00:00"},{"alias_kind":"pith_short_12","alias_value":"OMZRZRQAXJWP","created_at":"2026-05-20T00:04:34.372975+00:00"},{"alias_kind":"pith_short_16","alias_value":"OMZRZRQAXJWPO3XT","created_at":"2026-05-20T00:04:34.372975+00:00"},{"alias_kind":"pith_short_8","alias_value":"OMZRZRQA","created_at":"2026-05-20T00:04:34.372975+00:00"}],"events":[],"event_summary":{},"paper_claims":[],"inbound_citations":{"count":2,"internal_anchor_count":2,"sample":[{"citing_arxiv_id":"2605.18141","citing_title":"A Brief Overview: On-Policy Self-Distillation In Large Language Models","ref_index":18,"is_internal_anchor":true},{"citing_arxiv_id":"2605.18141","citing_title":"A Brief Overview: On-Policy Self-Distillation In Large Language Models","ref_index":18,"is_internal_anchor":true}]},"formal_canon":{"evidence_count":0,"sample":[],"anchors":[]},"links":{"html":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH","json":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH.json","graph_json":"https://pith.science/api/pith-number/OMZRZRQAXJWPO3XTE6QET63URH/graph.json","events_json":"https://pith.science/api/pith-number/OMZRZRQAXJWPO3XTE6QET63URH/events.json","paper":"https://pith.science/paper/OMZRZRQA"},"agent_actions":{"view_html":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH","download_json":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH.json","view_paper":"https://pith.science/paper/OMZRZRQA","resolve_alias":"https://pith.science/api/pith-number/resolve?arxiv=2605.05204&json=true","fetch_graph":"https://pith.science/api/pith-number/OMZRZRQAXJWPO3XTE6QET63URH/graph.json","fetch_events":"https://pith.science/api/pith-number/OMZRZRQAXJWPO3XTE6QET63URH/events.json","actions":{"anchor_timestamp":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH/action/timestamp_anchor","attest_storage":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH/action/storage_attestation","attest_author":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH/action/author_attestation","sign_citation":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH/action/citation_signature","submit_replication":"https://pith.science/pith/OMZRZRQAXJWPO3XTE6QET63URH/action/replication_record"}},"created_at":"2026-05-20T00:04:34.372975+00:00","updated_at":"2026-05-20T00:04:34.372975+00:00"}