ErwannMillon· @ErwannMillon · X·· 9 小时前AI 评分
Krea AI 分享 Flow Matching 模型调优经验
观察 Flow Matching 模型的去噪轨迹非常关键。深入理解模型在不同噪声水平下的行为，帮助我们在预训练中调整 timestep shift，获得了显著的效率提升。 这也帮助我们找到了感知上几乎无损的少步推理 schedule，为 timestep 蒸馏提供了最佳的初始化。
looking at the denoising trajectories of your flow matching model is so key. Deeply understanding what the model does at diff noise levels helped us tune the timestep shift during pretraining for big efficiency gains
It also helped us find the most perceptually lossless few-step inference schedule, which gave us the best possible initialization for timestep distillation
来源：ErwannMillon · x.com