You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
I was digging into the DiT paper and noticed something weird with the classifier-free guidance (CFG) equation. It seems like there might be a mistake compared to the original CFG formulation.
In the DiT paper, the equation is written as:
ε̂_θ(x_t, c) = ε_θ(x_t, ∅) + s * 🔴∇_x log p(x|c)
∝ ε_θ(x_t, ∅) + s * (ε_θ(x_t, c) - ε_θ(x_t, ∅))
But shouldn't it be:
ε̂_θ(x_t, c) = ε_θ(x_t, ∅) + s * 🔵∇_x log p(c|x)
∝ ε_θ(x_t, ∅) + s * (ε_θ(x_t, c) - ε_θ(x_t, ∅))
From what I understand, the guidance term should involve
∇_x log p(c|x), not ∇_x log p(x|c).
Can someone confirm if this is indeed an error, or am I missing something here?
Appreciate any insights!
I was digging into the DiT paper and noticed something weird with the classifier-free guidance (CFG) equation. It seems like there might be a mistake compared to the original CFG formulation.
In the DiT paper, the equation is written as:
But shouldn't it be:
From what I understand, the guidance term should involve
∇_x log p(c|x), not ∇_x log p(x|c).
Can someone confirm if this is indeed an error, or am I missing something here?
Appreciate any insights!
FYR.


✔ DiT
✔ CFG