models README

1
flow step t -> t_emb -> [MLP] -> [Linear] inside Ada -> split as γ,β
2
x = (1.0 + γ) * layer_norm(x) + β

1
s, t = sample_adjacent_times()                  # t < s. time=0 is noise => time=1 is clean.
2
x_t = add_noise(x0, noise(), t)
3
x_s = x_t + (s - t) * teacher(x_t, t)           # 教师做一步 Euler 求解
4
target = stop_grad(ema(x_s, s))
5
loss = distance(student(x_t, t), target)
6
optimize(student, loss); update_ema(ema, student)