PulseAugur
EN
LIVE 18:47:44

New theory explains and fixes instability in MeanFlow generative models

A new paper published on arXiv introduces a theoretical framework to address the instability issues encountered in MeanFlow training for generative models. The research identifies that the conditional velocity field is misused in the original MeanFlow loss, acting incorrectly as both a regression target and a control variate. The authors derive an optimal coefficient for the control variate role, unifying several concurrent remedies and demonstrating that this variance-optimal coefficient does not always align with the coefficient that yields the best generative quality. AI

IMPACT Provides theoretical grounding and practical fixes for unstable training in generative models, potentially improving their efficiency and quality.

RANK_REASON Academic paper detailing theoretical analysis and proposed solutions for a machine learning training method. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →

New theory explains and fixes instability in MeanFlow generative models

COVERAGE [1]

  1. arXiv cs.AI TIER_1 English(EN) · Juanwu Lu, Ziran Wang ·

    On Variance Reduction in Learning Mean Flows

    arXiv:2605.09235v2 Announce Type: replace-cross Abstract: One-step generative modeling has emerged as a leading approach for amortizing the inference cost of diffusion and flow-matching models. Among distillation-free methods, MeanFlow training is notoriously unstable, with non-d…