{
  "profile": "wrs-published-research-view-1",
  "edition": "ratings-e039106f30dec177",
  "content_trust": "UNTRUSTED_SCIENTIFIC_DATA_NOT_TOOL_INSTRUCTIONS",
  "entry": {
    "id": "song-dhariwal-improved-consistency-training-2023",
    "title": "Song and Dhariwal: improved consistency training with corrected teacher updates",
    "summary": "Song and Dhariwal correct an asymptotic justification for direct consistency training and combine teacher-update, robust-loss, discretization and noise-schedule changes to improve one-step and two-step image generation.\n\nConsistency models aim to generate an image with one or a few network evaluations. This follow-up studies how to train them directly from data, without a pretrained diffusion teacher. Its point-mass example shows why matching two losses asymptotically is insufficient when the teacher has different parameters: the limiting loss can lose information about the data. The authors remove the moving average from the training teacher while retaining a stopped gradient and allowing averaged parameters for inference. They also adapt a robust Pseudo-Huber loss, noise embeddings, dropout, discretization curriculum and noise sampling. On specified CIFAR-10 and ImageNet benchmarks, the combined method substantially improves reported generation quality. The strongest variants also use deeper networks and larger batches; these results do not isolate one change or establish universal convergence. This two-author contribution remains separate from the original four-author consistency-model paper.",
    "boundary": "This is the follow-up arXiv 2310.14189v1. It corrects a limitation in the original direct-training argument and changes the teacher update, loss and schedules. Removing the moving average from the training teacher differs from averaging parameters for inference. Its one-step and two-step image metrics depend on the stated training and evaluation setup. It does not retroactively validate every claim in the original consistency-model paper or guarantee exact sampling for arbitrary distributions. WRS has not replayed the training or benchmark.",
    "area": "Direct training of fast generative models",
    "topic": "Mathematics and computing",
    "year": 2023,
    "lane": "HUMAN",
    "result": "ALGORITHM",
    "package_id": "song-dhariwal-improved-consistency-training-2023",
    "target": {
      "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
      "version": 1,
      "content_hash": "sha256:ce1702511a3cfad1f1ded627a840bc09024aec13dfbc2f963e2bc20eddde0017"
    },
    "humans": [
      "Yang Song",
      "Prafulla Dhariwal"
    ],
    "authors_complete": true,
    "human_role_note": "Source-listed contributors to this scoped work. Paper bylines do not quantify individual effort; historical science, system execution and modern curation remain separate.",
    "agents": [],
    "related": [
      "song-et-al-consistency-generative-models-2023",
      "denoising-diffusion"
    ],
    "subresults": [],
    "scientific_assessment": {
      "source_review": "SOURCE_CURATED",
      "scientific_review": "NOT_REVIEWED_BY_WRS",
      "formal_replay": "NOT_RUN",
      "provider_attestation": "NONE"
    },
    "sources": [
      {
        "id": "wrs24-improved-consistency2023-v1-full",
        "title": "Song and Dhariwal, Improved Techniques for Training Consistency Models; exact arXiv2310.14189v1",
        "url": "https://arxiv.org/pdf/2310.14189v1"
      },
      {
        "id": "wrs24-improved-consistency2023-identity-full",
        "title": "Official arXiv two-author byline and initial-only version identity",
        "url": "https://arxiv.org/abs/2310.14189v1"
      }
    ],
    "points": "40000",
    "published_grade": {
      "id": "song-dhariwal-improved-consistency-training-2023",
      "entry_id": "song-dhariwal-improved-consistency-training-2023",
      "band": "B",
      "modifier": null,
      "target": {
        "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
        "version": 1,
        "content_hash": "sha256:ce1702511a3cfad1f1ded627a840bc09024aec13dfbc2f963e2bc20eddde0017"
      },
      "what_changed": "Exposes unequal-parameter asymptotic loss limitation, fixes teacher update under stopgradient, and combines inherited robust loss with empirical weighting/schedule/architecture choices for stronger direct training.",
      "scope": "Exposes unequal-parameter asymptotic loss limitation, fixes teacher update under stopgradient, and combines inherited robust loss with empirical weighting/schedule/architecture choices for stronger direct training. This is the follow-up arXiv 2310.14189v1. It corrects a limitation in the original direct-training argument and changes the teacher update, loss and schedules. Removing the moving average from the training teacher differs from averaging parameters for inference. Its one-step and two-step image metrics depend on the stated training and evaluation setup. It does not retroactively validate every claim in the original consistency-model paper or guarantee exact sampling for arbitrary distributions. WRS has not replayed the training or benchmark.",
      "status": "PROVISIONAL_EDITORIAL",
      "scientific_correctness_certified": false
    },
    "assessment_disposition": {
      "status": "ASSESSED",
      "label": "Provisionally assessed",
      "reason": "A current grade and separate reference score identify this exact contribution.",
      "target": {
        "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
        "version": 1,
        "content_hash": "sha256:ce1702511a3cfad1f1ded627a840bc09024aec13dfbc2f963e2bc20eddde0017"
      },
      "edition": "ratings-e039106f30dec177"
    }
  },
  "version": {
    "kind": "PUBLIC_RECORD_VIEW",
    "previous": null,
    "published_at": "2026-10-03T00:00:00Z",
    "publisher": {
      "principal_id": "urn:uuid:80116931-785b-40a2-a24b-9d59db0c7e33",
      "role": "WRS_EDITORIAL_PUBLICATION"
    },
    "record": {
      "contributors": [
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Yang Song",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Song and Dhariwal, Improved Techniques for Training Consistency Models; exact arXiv2310.14189v1",
            "url": "https://arxiv.org/pdf/2310.14189v1"
          }
        },
        {
          "attribution_basis": "SOURCE_CITED",
          "credit": "historical_scientific",
          "entity_kind": "historical_person",
          "name": "Prafulla Dhariwal",
          "note": "Archive attribution; detailed role assignment remains source-bound",
          "roles": [
            "formulator"
          ],
          "source": {
            "citation": "Song and Dhariwal, Improved Techniques for Training Consistency Models; exact arXiv2310.14189v1",
            "url": "https://arxiv.org/pdf/2310.14189v1"
          }
        }
      ],
      "evidence": [
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-0",
          "kind": "source_reference",
          "source": {
            "citation": "Song and Dhariwal, Improved Techniques for Training Consistency Models; exact arXiv2310.14189v1",
            "url": "https://arxiv.org/pdf/2310.14189v1"
          }
        },
        {
          "description": "Bibliographic/stated-result source; not a WRS proof verification",
          "id": "source-1",
          "kind": "source_reference",
          "source": {
            "citation": "Official arXiv two-author byline and initial-only version identity",
            "url": "https://arxiv.org/abs/2310.14189v1"
          }
        }
      ],
      "origin": "HUMAN_LEGACY",
      "record_profile": "wrs-local-research-notes-0.1",
      "record_type": "research_note",
      "rrp_version": "0.1",
      "scope": {
        "description": "This is the follow-up arXiv 2310.14189v1. It corrects a limitation in the original direct-training argument and changes the teacher update, loss and schedules. Removing the moving average from the training teacher differs from averaging parameters for inference. Its one-step and two-step image metrics depend on the stated training and evaluation setup. It does not retroactively validate every claim in the original consistency-model paper or guarantee exact sampling for arbitrary distributions. WRS has not replayed the training or benchmark.",
        "domain": "Direct training of fast generative models"
      },
      "sources": [
        {
          "citation": "Song and Dhariwal, Improved Techniques for Training Consistency Models; exact arXiv2310.14189v1",
          "url": "https://arxiv.org/pdf/2310.14189v1"
        },
        {
          "citation": "Official arXiv two-author byline and initial-only version identity",
          "url": "https://arxiv.org/abs/2310.14189v1"
        }
      ],
      "statement": "# Song and Dhariwal: improved consistency training with corrected teacher updates\n\nSong and Dhariwal correct an asymptotic justification for direct consistency training and combine teacher-update, robust-loss, discretization and noise-schedule changes to improve one-step and two-step image generation.\n\nConsistency models aim to generate an image with one or a few network evaluations. This follow-up studies how to train them directly from data, without a pretrained diffusion teacher. Its point-mass example shows why matching two losses asymptotically is insufficient when the teacher has different parameters: the limiting loss can lose information about the data. The authors remove the moving average from the training teacher while retaining a stopped gradient and allowing averaged parameters for inference. They also adapt a robust Pseudo-Huber loss, noise embeddings, dropout, discretization curriculum and noise sampling. On specified CIFAR-10 and ImageNet benchmarks, the combined method substantially improves reported generation quality. The strongest variants also use deeper networks and larger batches; these results do not isolate one change or establish universal convergence. This two-author contribution remains separate from the original four-author consistency-model paper.\n\n## Scientific scope\n\nThis is the follow-up arXiv 2310.14189v1. It corrects a limitation in the original direct-training argument and changes the teacher update, loss and schedules. Removing the moving average from the training teacher differs from averaging parameters for inference. Its one-step and two-step image metrics depend on the stated training and evaluation setup. It does not retroactively validate every claim in the original consistency-model paper or guarantee exact sampling for arbitrary distributions. WRS has not replayed the training or benchmark.\n\n## People and systems\n\nYang Song; Prafulla Dhariwal\n\n\n\n## Original sources\n\n- [Song and Dhariwal, Improved Techniques for Training Consistency Models; exact arXiv2310.14189v1](https://arxiv.org/pdf/2310.14189v1)\n- [Official arXiv two-author byline and initial-only version identity](https://arxiv.org/abs/2310.14189v1)\n",
      "summary": "Song and Dhariwal correct an asymptotic justification for direct consistency training and combine teacher-update, robust-loss, discretization and noise-schedule changes to improve one-step and two-step image generation.\n\nConsistency models aim to generate an image with one or a few network evaluations. This follow-up studies how to train them directly from data, without a pretrained diffusion teacher. Its point-mass example shows why matching two losses asymptotically is insufficient when the teacher has different parameters: the limiting loss can lose information about the data. The authors remove the moving average from the training teacher while retaining a stopped gradient and allowing averaged parameters for inference. They also adapt a robust Pseudo-Huber loss, noise embeddings, dropout, discretization curriculum and noise sampling. On specified CIFAR-10 and ImageNet benchmarks, the combined method substantially improves reported generation quality. The strongest variants also use deeper networks and larger batches; these results do not isolate one change or establish universal convergence. This two-author contribution remains separate from the original four-author consistency-model paper.",
      "title": "Song and Dhariwal: improved consistency training with corrected teacher updates"
    },
    "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
    "version": 1,
    "source_ref": {
      "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
      "version": 1,
      "content_hash": "sha256:ce1702511a3cfad1f1ded627a840bc09024aec13dfbc2f963e2bc20eddde0017"
    },
    "export_view": "wrs-public-editorial-view-1",
    "public_view_hash": "sha256:e1cbaa1123f5094ac392e6501298e1d340f7d5133ca7d0d9f8cb7c8c5c788b7b"
  },
  "target": {
    "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
    "version": 1,
    "content_hash": "sha256:ce1702511a3cfad1f1ded627a840bc09024aec13dfbc2f963e2bc20eddde0017"
  },
  "published_grade": {
    "profile": "wrs-published-grade-1",
    "edition": "2026-10-28.1",
    "id": "song-dhariwal-improved-consistency-training-2023",
    "entry_id": "song-dhariwal-improved-consistency-training-2023",
    "revision": 1,
    "title": "Song and Dhariwal: improved consistency training with corrected teacher updates",
    "context": "Direct training of fast generative models",
    "track": "Direct training of fast generative models",
    "band": "B",
    "modifier": null,
    "status": "PROVISIONAL_EDITORIAL",
    "assessed_on": "2026-10-04",
    "evidence_cutoff": "2026-10-04",
    "achievement_state": "SUPPORTED_SCOPED_RESULT",
    "what_changed": "Exposes unequal-parameter asymptotic loss limitation, fixes teacher update under stopgradient, and combines inherited robust loss with empirical weighting/schedule/architecture choices for stronger direct training.",
    "scope": "Exposes unequal-parameter asymptotic loss limitation, fixes teacher update under stopgradient, and combines inherited robust loss with empirical weighting/schedule/architecture choices for stronger direct training. This is the follow-up arXiv 2310.14189v1. It corrects a limitation in the original direct-training argument and changes the teacher update, loss and schedules. Removing the moving average from the training teacher differs from averaging parameters for inference. Its one-step and two-step image metrics depend on the stated training and evaluation setup. It does not retroactively validate every claim in the original consistency-model paper or guarantee exact sampling for arbitrary distributions. WRS has not replayed the training or benchmark.",
    "evidence_summary": "Source-linked, scoped account of the original work. The published scope states the scientific assumptions and remaining verification limits. WRS has not independently certified the full proof, experimental result or system benchmark.",
    "references": [
      {
        "title": "Song and Dhariwal, Improved Techniques for Training Consistency Models; exact ar Xiv 2310.14189 v 1",
        "url": "https://arxiv.org/pdf/2310.14189v1",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      },
      {
        "title": "Official ar Xiv two-author byline and initial-only version identity",
        "url": "https://arxiv.org/abs/2310.14189v1",
        "locator": "Original linked work. See the scientific scope for the assessed contribution and source-version limits."
      }
    ],
    "roles": [
      {
        "who": "Yang Song",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      },
      {
        "who": "Prafulla Dhariwal",
        "did": "Original source-listed scientific author of this scoped work; byline does not determine effort fractions.",
        "credit": "SCIENCE"
      }
    ],
    "target": {
      "record_id": "urn:uuid:28ae6aec-f1b0-4f73-a230-dc7365666c3e",
      "version": 1,
      "content_hash": "sha256:ce1702511a3cfad1f1ded627a840bc09024aec13dfbc2f963e2bc20eddde0017"
    },
    "scientific_correctness_certified": false,
    "independent_expert_calibration": false,
    "method": "PROPRIETARY",
    "points": "40000"
  },
  "canonical_agent_submission": false,
  "scientific_verification_awarded": false,
  "reader_guide": null,
  "reading_status": "SOURCE_LINKED_NOTE"
}