parsee-mizuhashi
commited on
Commit
•
80949ae
1
Parent(s):
6975069
Upload 2 files
Browse files- index.html +92 -17
- style.css +123 -18
index.html
CHANGED
@@ -1,19 +1,94 @@
|
|
1 |
<!DOCTYPE html>
|
2 |
-
<html>
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
|
15 |
-
|
16 |
-
|
17 |
-
|
18 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
19 |
</html>
|
|
|
1 |
<!DOCTYPE html>
|
2 |
+
<html lang="en">
|
3 |
+
<head>
|
4 |
+
<meta charset="UTF-8">
|
5 |
+
<meta name="viewport" content="width=device-width, initial-scale=1.0">
|
6 |
+
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.3/css/all.min.css">
|
7 |
+
<link href="https://fonts.googleapis.com/css2?family=Montserrat:wght@500&display=swap" rel="stylesheet">
|
8 |
+
<link rel="stylesheet" href="style.css">
|
9 |
+
<title>M.o.f.u.</title>
|
10 |
+
</head>
|
11 |
+
<body>
|
12 |
+
<h1 class="header-title">M.o.f.u.</h1>
|
13 |
+
<p class="header-subtitle"><span class="highlight-orange">Mo</span>del independent, <span class="highlight-violate">F</span>ast T<span class="highlight-orange">u</span>ning of Stable Diffusion concepts</p>
|
14 |
+
<section id="abstract">
|
15 |
+
<h2><i class="icon fas fa-file-alt"></i> Abstract</h2>
|
16 |
+
<p>I present MoFu, a model-independent, fast tuning
|
17 |
+
approach that enhances Stable Diffusion. Compared
|
18 |
+
to other more traditional methods, such as Low Rank
|
19 |
+
adaptation for the model or fine tuning it, MoFu
|
20 |
+
doesn’t modify the weights of the main model at all.
|
21 |
+
MoFu seamlessly integrates with Stable Diffusion's
|
22 |
+
text encoder, enabling rapid style/concept addition
|
23 |
+
without modifying or fine-tuning the encoder's
|
24 |
+
weights</p>
|
25 |
+
</section>
|
26 |
+
|
27 |
+
<section id="methodology">
|
28 |
+
<h2><i class="icon fas fa-flask"></i> Methodology</h2>
|
29 |
+
<p>The methodology of MoFu revolves around a simple
|
30 |
+
yet effective process. To achieve the desired results,
|
31 |
+
we begin by comparing natural prompts given to a
|
32 |
+
set of images. This comparison allows us to extract
|
33 |
+
the essential concepts or styles from the text
|
34 |
+
prompts. These identified concepts are then stored in
|
35 |
+
a mixin, creating a compact representation of the
|
36 |
+
desired style information. The mixin is designed to be
|
37 |
+
compatible with Stable Diffusion's architecture and
|
38 |
+
serves as an additive to the text encoder’s output.
|
39 |
+
By adding the mixin with the text encoder’s output
|
40 |
+
(the mixin, or MoFu model, can also be multiplied by
|
41 |
+
a weight, in order to make its effect stronger or
|
42 |
+
weaker), MoFu efficiently injects the extracted
|
43 |
+
concepts into the image generation process. This
|
44 |
+
injection enables Stable Diffusion to generate images
|
45 |
+
with the desired style without altering the underlying
|
46 |
+
weights of the main model. As a result, MoFu
|
47 |
+
provides a powerful and flexible solution for style
|
48 |
+
transfer or concept addition in Stable Diffusion
|
49 |
+
without the need for extensive model modifications
|
50 |
+
or resource-intensive fine-tuning.</p>
|
51 |
+
</section>
|
52 |
+
|
53 |
+
<section id="results">
|
54 |
+
<h2><i class="icon fas fa-chart-bar"></i> Results</h2>
|
55 |
+
<p>To evaluate the effectiveness of MoFu, I conducted a
|
56 |
+
series of experiments and compared its performance
|
57 |
+
to LoRA and fine-tuning methods. Our results
|
58 |
+
demonstrate that MoFu achieves comparable
|
59 |
+
performance to LoRAs while requiring significantly
|
60 |
+
less training time, taking only around 10-20 seconds
|
61 |
+
on average, primarily due to being CPU-bound. This
|
62 |
+
is in stark contrast to LoRAs, which typically demand
|
63 |
+
several hours to train. However, I also observed that
|
64 |
+
MoFu falls short of fine-tuning, as the latter can
|
65 |
+
achieve even better precision/quality but at the cost
|
66 |
+
of a much longer training.
|
67 |
+
</p>
|
68 |
+
</section>
|
69 |
+
|
70 |
+
<section id="conclusion">
|
71 |
+
<h2><i class="icon fas fa-clipboard-check"></i> Conclusion</h2>
|
72 |
+
<p>In conclusion, MoFu offers an efficient and
|
73 |
+
model-independent solution for adding new styles or
|
74 |
+
concepts to Stable Diffusion without modifying the
|
75 |
+
main model's weights. It achieves comparable results
|
76 |
+
to LoRA while significantly reducing training time,
|
77 |
+
making it a practical choice for rapid adaptation.
|
78 |
+
Though fine-tuning still outperforms MoFu in quality,
|
79 |
+
the trade-off between speed and accuracy makes
|
80 |
+
MoFu a valuable option for various applications.
|
81 |
+
Future work may focus on optimizing the
|
82 |
+
implementation / quality of MoFu.</p>
|
83 |
+
</section>
|
84 |
+
|
85 |
+
<footer>
|
86 |
+
<div class="buttons">
|
87 |
+
<a href="mailto:parsee.mizuhashi.th11@gmail.com" class="button">Yoinked</a>
|
88 |
+
<a href="https://huggingface.co/organizations/touhou-ai-experimental" class="button">Touhou AI Experimental Group</a>
|
89 |
+
<a href="https://huggingface.co/mofu-team" class="button">MoFu</a>
|
90 |
+
<a href="https://github.com/yoinked-h/MoFu" class="button" target="_blank">GitHub</a>
|
91 |
+
</div>
|
92 |
+
</footer>
|
93 |
+
</body>
|
94 |
</html>
|
style.css
CHANGED
@@ -1,28 +1,133 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
body {
|
2 |
-
|
3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
4 |
}
|
5 |
|
6 |
-
|
7 |
-
|
8 |
-
|
|
|
|
|
|
|
9 |
}
|
10 |
|
11 |
-
|
12 |
-
|
13 |
-
font-size: 15px;
|
14 |
-
margin-bottom: 10px;
|
15 |
-
margin-top: 5px;
|
16 |
}
|
17 |
|
18 |
-
|
19 |
-
|
20 |
-
|
21 |
-
padding: 16px;
|
22 |
-
border: 1px solid lightgray;
|
23 |
-
border-radius: 16px;
|
24 |
}
|
25 |
|
26 |
-
.
|
27 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
28 |
}
|
|
|
1 |
+
/* Reset some default styling */
|
2 |
+
body, h1, h2, p {
|
3 |
+
margin: 0;
|
4 |
+
padding: 0;
|
5 |
+
}
|
6 |
+
|
7 |
body {
|
8 |
+
font-family: Arial, sans-serif;
|
9 |
+
background-color: #121212;
|
10 |
+
color: #FFFFFF;
|
11 |
+
}
|
12 |
+
|
13 |
+
.header-title {
|
14 |
+
font-family: 'Montserrat', sans-serif;
|
15 |
+
font-size: 24px;
|
16 |
+
text-align: center;
|
17 |
+
color: #FF4500; /* Original Orange-Red Color */
|
18 |
+
margin: 0;
|
19 |
+
padding: 20px 0;
|
20 |
+
}
|
21 |
+
.header-subtitle {
|
22 |
+
font-size: 16px;
|
23 |
+
text-align: center;
|
24 |
+
color: #FF4500;
|
25 |
+
margin: 10px 0;
|
26 |
+
}
|
27 |
+
|
28 |
+
.highlight-orange {
|
29 |
+
color: #EE82EE; /* Original Orange-Red Color */
|
30 |
+
}
|
31 |
+
|
32 |
+
.highlight-violate {
|
33 |
+
color: #EE82EE; /* Violate Color */
|
34 |
+
}
|
35 |
+
header {
|
36 |
+
display: flex;
|
37 |
+
justify-content: space-between;
|
38 |
+
align-items: center;
|
39 |
+
background-color: #FF4500; /* Original Orange-Red Color */
|
40 |
+
padding: 10px 20px;
|
41 |
+
}
|
42 |
+
|
43 |
+
nav ul {
|
44 |
+
list-style-type: none;
|
45 |
+
display: flex;
|
46 |
+
}
|
47 |
+
|
48 |
+
nav li {
|
49 |
+
margin-right: 20px;
|
50 |
+
}
|
51 |
+
|
52 |
+
nav a {
|
53 |
+
text-decoration: none;
|
54 |
+
color: #FFFFFF;
|
55 |
}
|
56 |
|
57 |
+
section {
|
58 |
+
background-color: #1E1E1E;
|
59 |
+
border: 2px solid #FF4500;
|
60 |
+
border-radius: 10px;
|
61 |
+
padding: 20px;
|
62 |
+
margin: 20px;
|
63 |
}
|
64 |
|
65 |
+
section:hover {
|
66 |
+
border-color: #EE82EE;
|
|
|
|
|
|
|
67 |
}
|
68 |
|
69 |
+
h2 {
|
70 |
+
color: #FF4500;
|
71 |
+
margin-bottom: 10px;
|
|
|
|
|
|
|
72 |
}
|
73 |
|
74 |
+
.icon {
|
75 |
+
margin-right: 10px;
|
76 |
+
}
|
77 |
+
|
78 |
+
.footer {
|
79 |
+
text-align: center;
|
80 |
+
margin-top: 40px;
|
81 |
+
}
|
82 |
+
|
83 |
+
.buttons {
|
84 |
+
display: flex;
|
85 |
+
justify-content: center;
|
86 |
+
align-items: center;
|
87 |
+
margin: 20px auto;
|
88 |
+
}
|
89 |
+
|
90 |
+
.button {
|
91 |
+
background-color: #FF4500;
|
92 |
+
color: #121212;
|
93 |
+
padding: 8px 16px;
|
94 |
+
border-radius: 10px;
|
95 |
+
text-decoration: none;
|
96 |
+
font-weight: bold;
|
97 |
+
transition: background-color 0.3s;
|
98 |
+
margin-right: 10px;
|
99 |
+
}
|
100 |
+
|
101 |
+
.button:last-child {
|
102 |
+
margin-right: 0;
|
103 |
+
}
|
104 |
+
|
105 |
+
.button:hover {
|
106 |
+
background-color: #EE82EE;
|
107 |
+
}
|
108 |
+
|
109 |
+
/* Media Queries for Responsive Design */
|
110 |
+
@media (max-width: 768px) {
|
111 |
+
header {
|
112 |
+
padding: 10px;
|
113 |
+
}
|
114 |
+
|
115 |
+
nav ul {
|
116 |
+
flex-direction: row;
|
117 |
+
}
|
118 |
+
|
119 |
+
nav li {
|
120 |
+
margin-right: 10px;
|
121 |
+
}
|
122 |
+
|
123 |
+
.buttons {
|
124 |
+
flex-direction: column;
|
125 |
+
align-items: center;
|
126 |
+
margin: 10px auto;
|
127 |
+
}
|
128 |
+
|
129 |
+
.button {
|
130 |
+
margin-right: 0;
|
131 |
+
margin-bottom: 10px;
|
132 |
+
}
|
133 |
}
|