Mean Flows for One-step Generative Modeling
Zhengyang Geng, Mingyang Deng, Xingjian Bai, J. Zico Kolter et al.
71.43/100
📘 Readable
Decent, has merit
Content 62.4 · Citation bonus +9.0 · 425 citations
💡 This paper proposes MeanFlow, which replaces instantaneous velocity in Flow Matching with average velocity, derives a velocity identity for training, and achieves 3.43 FID on ImageNet 256×256 with sin
#一步出图#Flow Matchin#平均速度真香#推理快如闪电#生成模型爽文#one-shot gen#Flow Matching revamp#avg velocity hype#lightning inference#gen model爽文
Score breakdown
Novelty8.0 / 10
Rigor7.0 / 10
Significance9.0 / 10
Clarity8.0 / 10
Reproducibility7.0 / 10
This tone hasn't been generated yet — roast it again to create it.