-
Notifications
You must be signed in to change notification settings - Fork 0
/
index.html
360 lines (323 loc) · 17.6 KB
/
index.html
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
<html>
<head lang="en">
<!-- https://stackoverflow.com/questions/16316311/github-pages-and-relative-paths -->
<base href="/[repo]/">
<meta charset="UTF-8">
<meta http-equiv="x-ua-compatible" content="ie=edge">
<title>MultiFusion: Fusing Pre-Trained Models for Multi-Lingual, Multi-Modal Image Generation</title>
<meta name="description" content="">
<meta name="viewport" content="width=device-width, initial-scale=1">
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.5/css/bootstrap.min.css">
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/font-awesome/4.4.0/css/font-awesome.min.css">
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/codemirror/5.8.0/codemirror.min.css">
<link rel="stylesheet" href="css/app.css">
<link rel="stylesheet" href="css/bootstrap.min.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/1.11.3/jquery.min.js"></script>
<script src="https://maxcdn.bootstrapcdn.com/bootstrap/3.3.5/js/bootstrap.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/codemirror/5.8.0/codemirror.min.js"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/clipboard.js/1.5.3/clipboard.min.js"></script>
<script src="js/app.js"></script>
</head>
<body>
<div class="container" id="main" style="background-color: rgba(255, 255, 255, 0.9);">
<div class="row">
<h2 class="col-md-12 text-center">
MultiFusion: Fusing Pre-Trained Models for Multi-Lingual, Multi-Modal Image Generation
</h2>
</div>
<style>
.img_container {
display: flex;
align-items: center;
justify-content: center;
}
.author-info {
text-align: center;
margin-right: 40px; /* Adjust this to move the block to the right */
}
.author {
display: flex;
justify-content: center;
gap: 35px; /* Change this value to adjust space between authors */
}
.author li, .affiliation {
list-style-type: none; /* This will remove the bullet points */
}
.affiliation {
margin-top: 5px; /* Smaller values to reduce space */
margin-bottom: 5px;
}
.author-name {
font-size: 20px; /* Adjust this value to increase or decrease the author name size */
}
.author-aff {
font-size: 10px; /* Adjust this value to increase or decrease the author name size */
color:#636666;
}
/* Medium screens (Tablets) */
@media only screen and (max-width: 1024px) {
.author-name {
font-size: 18px;
}
.affiliation {
font-size: 13px;
}
.author {
gap: 20px;
}
.author-aff {
font-size: 9px;
}
}
/* Medium screens (Tablets) */
@media only screen and (max-width: 768px) {
.author-name {
font-size: 16px;
}
.affiliation {
font-size: 12px;
}
.author {
gap: 15px;
}
.author-aff {
font-size: 8px;
}
}
</style>
<div class="row">
<div class="col-md-12 text-center">
<ul class="author-info">
<div class="author">
<li>
<p class="author-name">Marco Bellagente</p><sup class="author-aff">*+ 4</sup><br/>
</li>
<li>
<p class="author-name">Manuel Brack</p><sup class="author-aff">* 2,3</sup><br/>
</li>
<li>
<p class="author-name">Hannah Teufel</p><sup class="author-aff">* 1</sup><br/>
</li>
</div>
<div class="author">
<li>
<p class="author-name">Felix Friedrich</p><sup class="author-aff">3,6</sup><br/>
</li>
<li>
<p class="author-name">Björn Deiseroth</p><sup class="author-aff">1,3,6</sup><br/>
</li>
<li>
<p class="author-name">Constantin Eichenberg</p><sup class="author-aff">1</sup><br/>
</li>
<li>
<p class="author-name">Andrew Dai</p><sup class="author-aff">1</sup><br/>
</li>
<li>
<p class="author-name">Robert J.N. Baldock</p><sup class="author-aff">1</sup><br/>
</li>
</div>
<div class="author">
<li>
<p class="author-name">Souradeep Nanda</p><sup class="author-aff">+ 5</sup><br/>
</li>
<li>
<p class="author-name">Koen Oostermeijer</p><sup class="author-aff">1</sup><br/>
</li>
<li>
<p class="author-name">Andres Felipe Cruz-Salinas</p><sup class="author-aff">1</sup><br/>
</li>
<li>
<p class="author-name">Patrich Schramowski</p><sup class="author-aff">2,3,6,8</sup><br/>
</li>
</div>
<div class="author">
<li>
<p class="author-name">Kristian Kersting</p><sup class="author-aff">✦ 2,3,6,7</sup><br/>
</li>
<li>
<p class="author-name">Samuel Weinbach</p><sup class="author-aff">✦ 1</sup><br/>
</li>
</div>
<li class="affiliation">
<sup>1</sup>Aleph Alpha   <sup>2</sup>German Research Center for Artificial Intelligence (DFKI)
</li>
<li class="affiliation">
<sup>3</sup>Computer Science Department, TU Darmstadt   <sup>4</sup>Stability AI   <sup>5</sup>University of Texas
</li>
<li class="affiliation">
<sup>6</sup>Hessian.AI   <sup>7</sup>Centre for Cognitive Science, TU Darmstadt   <sup>8</sup>LAION  
</li>
<li class="affiliation">
(+) Work performed while at Aleph Alpha
(*) equal contribution
(✦) equal supervision
</li>
</ul>
</div>
</div>
<style>
.link-content {
text-align: center;
}
.list-inline li {
margin-right: 90px; /* Adjust this value to increase/decrease the gap between items */
margin-left: 90px;
}
</style>
<div class="row">
<div class="col-md-12 text-center">
<ul class="list-inline">
<li>
<a href="https://arxiv.org/abs/2305.15296">
<div class="link-content">
<h4><strong>Paper</strong></h4>
</div>
</a>
</li>
<li>
<a href="https://huggingface.co/datasets/AIML-TUDA/MCC-250">
<div class="link-content">
<h4><strong>MCC-250 Benchmark Dataset</strong></h4>
</div>
</a>
</li>
</ul>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Abstract
</h3>
<p class="text-justify">
The recent popularity of text-to-image diffusion models (DM) can largely be attributed to
the intuitive interface they provide to users. The intended generation can be expressed in natural language, with the model producing faithful interpretations of text prompts.
However, expressing complex or nuanced ideas in text alone can be difficult. To ease image generation,
we propose MultiFusion that allows one to express complex and nuanced concepts with arbitrarily interleaved inputs of multiple modalities and languages.
MultiFusion leverages pre-trained models and aligns them for integration into a cohesive system, thereby avoiding the need for extensive training from scratch.
Our experimental results demonstrate the efficient transfer of capabilities from individual modules to the downstream model. Specifically, the fusion of all independent components allows the image generation module to utilize multilingual, interleaved multimodal inputs despite being trained solely on monomodal data in a single language.
</p>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Method
</h3>
<p class="text-justify">
To enable multimodal, multilingual prompting, both compute efficiently and without multimodal downstream training data we use a custom modular encoder.
The base of our encoder is a 13B autoregressive transformer (1.1), pretrained on 5 languages (English, German, Spanish, Italian and French).
We extend the encoder by an image prefix as well as adapters (1.2) to enable multimodality. Additionally, we finetune the biases (2.1) of the LLM to provide embeddings, which capture the semantic meaning of
the text prompt, thus simplifying the learning of mapping from embeddings to image outputs. Finally, to align the pre-trained Stable Diffusion model (1.4) with the embeddings of our modular
encoder, we retrain the conditioning by finetuning the cross-attention weights (2.2).
</p>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/architecture.png" class="img-responsive" alt="architecture figure 5"><br>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Evaluation
</h3>
<p>In the following section we provide a concise overview of the quantitative and qualitative evaluation of MultiFusion.</p>
<h4>Image Fidelity and Text-to-Image Alignment</h4>
<p>First we meassure image fidelity and image-text-alignment using the standard metrics FID-30K and Clip Scores. We find that MultiFusion prompted with text only performs on par with Stable Diffusion despite extension of the Encoder to support multiple languages and modalities.</p>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/evaluation.png" class="img-responsive"><br>
<h4>Compositional Robustness</h4>
<div class="img_container" style="margin-bottom:20px">
<div float="left">
<image height="150px" src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/compositional_robustness.png" alt="method"><br>
</div>
<div float="left" style="margin-bottom:20px">
<p>Image Composition is a known limitation of Diffusion Models. Through evaluation of our new benchmark <a href="https://huggingface.co/datasets/AIML-TUDA/MCC-250">MCC-250</a> we show that multimodal prompting leads to more compositional robustness as judged by humans. Each prompt is a complex conjunction of two different objects with different
colors, with multimodal prompts containing one visual reference for each object interleaved with the text input. </p>
</div>
</div>
<h4>Multilinguality</h4>
<p>Below we demostrate the multilingual alignment of images generated by MultiFusion. All images were generated using the same seed and with the respective translation of the prompt ‘an image of an astronaut riding
a horse’.</p>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/multilingual_alignment.png" class="img-responsive" alt="method"><br>
<p> We show the comparison of multilingual alignment over DrawBench prompts. MultiFusion achieves
comparable alignment of the output images although the image generation module was only
trained on English data. This can be attributed to the strong alignment of multilingual prompts
in MultiFusion’s embedding space. </p>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/multilingual_evaluation_graphs.png" class="img-responsive" alt="method"><br>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Attention Manipulation for Multimodal Inference
</h3>
<p> Attention Manipulation, based on <a href="https://arxiv.org/abs/2301.08110">AtMan</a>, allows us to weight image and text tokens at inference time and guide their influence on the resulting generation. </p>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/attn_manipulation.png" class="img-responsive" alt="method"><br>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Applications
</h3>
<p>Finally, we present use cases and applications demonstrating the unique capabilities of MultiFusion.</p>
<h4>Interleaved multilingual, multimodal prompting</h4>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/prompting_example.png" class="img-responsive" alt="method"><br>
<div class="img_container" style="margin-bottom:20px">
<div>
<h4>Image Composition</h4>
<p>MultiFusion increases expressivness in composition through arbitrary and flexible promptin of image and text sequences.</p>
<image height="130px" src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/image_composition.png" alt="method"><br>
</div>
<div>
<h4>Negative Prompting</h4>
<p>Negative prompting with images enables a more powerful supression than through text prompts.</p>
<image height="130px" src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/negative_prompting.png" alt="method"><br>
</div>
</div>
<h4>Style Modification</h4>
MultiFusion enables simple style transfer through one reference image capturing all the facets of a unique style such as color pallette, composition contrast, etc. making elaborate prompts obsolete. Additionally, MultiFusion enables highly individual prompting such as "in the style of a picture I drew".
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/fig5e_boxed.png" class="img-responsive" alt="method"><br>
<h4>Image Variation</h4>
<p>MultiFusion produces meaningful image variations without the need for inversion or renoising if the input image.</p>
<image src="https://Aleph-Alpha.github.io/MultiFusion/src/imgs/fig5d_boxed.png" class="img-responsive" alt="method"><br>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Follow-Up Work
</h3>
<p>
In order to contribute to the fair evaluation and safety of large generative models we opened our model for the evaluation on the Holisitic Evaluation of Text-to-Image Models (HEIM) benchmark [<a href="https://crfm.stanford.edu/heim/v1.1.0/">results</a>, <a href="https://arxiv.org/abs/2311.04287">paper</a>]. As well as for a study on the mitigation of inapproriatness in image generation [<a href="https://arxiv.org/abs/2305.18398">paper</a>]
</p>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Citation
</h3>
<div class="form-group col-md-10 col-md-offset-1">
<pre id="bibtex" readonly>
@article{bellagente2023multifusion,
title={MultiFusion: Fusing Pre-Trained Models for Multi-Lingual, Multi-Modal Image Generation},
author={Marco Bellagente and Manuel Brack and Hannah Teufel and Felix Friedrich and Björn Deiseroth and Constantin Eichenberg and Andrew Dai and Robert Baldock and Souradeep Nanda and Koen Oostermeijer and Andres Felipe Cruz-Salinas and Patrick Schramowski and Kristian Kersting and Samuel Weinbach},
year={2023},
journal={arXiv preprint arXiv:2305.15296},
}</pre>
</div>
</div>
</div>
<div class="row">
<div class="col-md-8 col-md-offset-2">
<h3>
Acknowledgements
</h3>
<p class="text-justify">
The website template was borrowed from <a href="https://jonbarron.info/">Jon Barron</a>.
</p>
<br>
</div>
</div>
</div>
</body>
</html>