BWhat 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.
Scientific 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
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.
Contributors
- Yang Song · SCIENCE: Original source-listed scientific author of this scoped work; byline does not determine effort fractions.
- Prafulla Dhariwal · SCIENCE: Original source-listed scientific author of this scoped work; byline does not determine effort fractions.
- Song and Dhariwal, Improved Techniques for Training Consistency Models; exact ar Xiv 2310.14189 v 1
- Official ar Xiv two-author byline and initial-only version identity
Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction