TBSM 提出一步生成模型训练新法：轻量 Tracker 替代复杂蒸馏
TBSM 用轻量 Tracker 指导生成样本走向真实分布，实现稳定的一步生成，在 ImageNet 上达到较低 FID。
TBSM 训练一个不足 30M 参数的 Tracker，在线学习生成样本应向真实分布移动的方向。生成器只需回归移动后的位置，即可用一次前向生成图像。在 ImageNet-256 上，潜在空间 FID 1.63，像素空间 2.23，推理时仅需一次函数评估。
正文摘录
Title: How to Train One-Step Generative Models Without CFG, DMD, GAN, Drifting, and MeanFlow? Source: 机器之心 (Machine Intelligence) --- .jpg)  Generating an image with just a single forward pass through the generator – that's the efficiency that one-step generation pursues. But to reduce the sampling steps to 1, the real challenge lies on the training side: how do you use a sufficiently stable and simple supervisory signal to make the generated distribution approach the real one? Existing approaches each come w…