QGF's two cuts · peeling back the guidance gradient ring by ring

Horizontal axis = noisy point a_t, vertical axis = guidance gradient. Click ①②③ to peel back ring by ring and watch the gradient grow smooth.

① BPTT (full chain) ② one step + exact J ③ QGF (J=I)

time t 0.60

probe a_t 0.45

① BPTT jitter |Δg|

—

② one step + J jitter

—

③ QGF jitter

—