如上图,训练时候的文本提示
c
c
c用空提示(一个固定的概率)
c
ϕ
c_{\phi}
cϕ代替;在推理阶段,模型推断在guidance scale
s
>
=
1
s>=1
s>=1之下,朝着
ϵ
θ
(
z
t
,
t
,
c
)
\epsilon_{\theta}(z_t,t,c)
ϵθ(zt,t,c)的方向,远离
ϵ
θ
(
z
t
,
t
,
c
ϕ
)
\epsilon_{\theta}(z_t,t,c_{\phi})
ϵθ(zt,t,cϕ)
借鉴LaMa的mask方法:irregular masks (thick, medium, and thin masks) ,which uniformly uses polygonal chains dilated by a high random width (wide masks) and rectangles of arbitrary aspect ratios (box masks).