MM
-
DiT
Blocks
×
10
Single
-
DiT
Blocks
×
20
Text
Prompt
VAE
T5 text encoder
Glyph Byt5
VAE
N
×
MMDIT Blocks
Norm + Out Projection
…
…
…
…
…
…
Patchify
RF Loss
RF Loss
TEncoder
RTP Loss
Unpatchify
+ Decode
LoRA
!
(a) Training Framework
(c)
LoRA
Control
(b) Region Perceptual loss
norm
MLP
MM
Attention
Out Projection
⨂
⨁
norm
MLP
MM
Attention
Out Projection
⨂
⨁
LoRA
X
!
X
"
X
"
X
!
X
#
Word
Prompt
"
"$%&
"
'
X
'
M
-
RoPE
🔥
🔥
𝑡
!
𝑡
#
𝑡
$
𝑡
%
&
#
…
𝑡
%
𝑣
#
$
𝑣
#
#
𝑣
#
'
𝑣
(
)
&
#
𝑣
(
)
卡通扁平风格的餐饮招牌推荐图,以鲜艳红色为背
景,左上角有醒目的“
”艺术字,搭配黄
色点赞手势图标,右侧是装满肉类、蔬菜、菌类等各
类食材的彩色火锅插画,背景中有白色星星装饰,下
方是拼音“
”
。整
体布局干净简洁,充满餐厅促销物料的活力感。
招
牌
推
荐
Z
H
A
O
P
A
I
T
U
I
J
I
A
N
0,
0,
0,
0,
1,
1,
0,
n,
n
0,
2,
2,
0,
n
-
1,
n
-
1,
...
1,
n+1,
n+1,
1,
n+1,
n+2,
1,
n+1,
n+3,
1,
n+h
,
n+w
-
1,
1,
n+h
,
n+w
…
VAE
Encoder
❄
Qwen2.5
VL
❄
…
…
Timestep
Attention
Q
K
V
Q
K
V
Attn
-
out
Attn
-
out
MLP
-
in/out
MLP
-
in/out
Attention
MLP
-
in
Q
K
V
⨁
…
…
MLP
-
out
Attn
-
out
…