-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathindex.html
574 lines (513 loc) · 22.8 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
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
<!doctype html>
<html lang="en">
<head>
<!-- Required meta tags -->
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1, shrink-to-fit=no">
<link rel="icon" href="path/to/favicon_F.ico" type="./src/img/favicon.ico">
<!-- Bootstrap CSS -->
<link rel="stylesheet" href="https://maxcdn.bootstrapcdn.com/bootstrap/4.0.0/css/bootstrap.min.css"
integrity="sha384-Gn5384xqQ1aoWXA+058RXPxPg6fy4IWvTNh0E263XmFcJlSAwiGgFAW/dAiS6JXm" crossorigin="anonymous">
<link href="https://fonts.googleapis.com/css2?family=Roboto:wght@300&display=swap" rel="stylesheet">
<title>FLOAT</title>
<style>
*{
font-family: 'Roboto', sans-serif;
padding: 0;
margin: 0;
outline: none;
}
.video-container {
display: flex;
gap: 10px;
justify-content: center;
}
.video-container video {
width: 100%;
max-width: 365px;
height: auto;
}
.embed-responsive {
position: relative;
overflow: hidden;
}
.embed-responsive-item {
position: absolute;
top: 0;
left: 0;
width: 100%;
height: 100%;
object-fit: contain;
}
video {
max-width: 100%;
height: auto;
display: block;
}
.btn-primary {
background-color: navy;
border-color: navy;
color: white;
}
.btn-primary:hover {
background-color: #004080;
border-color: #004080;
}
a {
color: navy;
text-decoration: none;
padding: 5px 10px;
border-radius: 5px;
}
a:hover {
color: #004080;
}
.custom-control-prev {
position: absolute;
left: -250px;
top: 50%;
transform: translateY(-50%);
z-index: 5;
}
.custom-control-next {
position: absolute;
right: -250px;
top: 50%;
transform: translateY(-50%);
z-index: 5;
}
.carousel-item video {
max-width: 100%;
height: auto;
}
.carousel-control-prev {
position: absolute;
left: -50px;
top: 50%;
transform: translateY(-50%);
z-index: 10;
font-size: 30px;
width: 60px;
height: 60px;
background-color: rgba(0, 0, 0, 0.5);
color: white;
border-radius: 50%;
display: flex;
justify-content: center;
align-items: center;
}
.carousel-control-next {
position: absolute;
right: -50px;
top: 50%;
transform: translateY(-50%);
z-index: 10;
font-size: 30px;
width: 60px;
height: 60px;
background-color: rgba(0, 0, 0, 0.5);
color: white;
border-radius: 50%;
display: flex;
justify-content: center;
align-items: center;
}
.video-fit {
width: 100%;
height: auto;
max-height: calc(100vh - 100px);
object-fit: contain;
}
.carousel-control-prev:hover, .carousel-control-next:hover {
background-color: rgba(0, 0, 0, 0.8);
}
</style>
</head>
<body>
<div class="container">
<br>
<div style="text-align: center;">
<h1>FLOAT</h1>
<h3> Generative Motion Latent Flow Matching for Audio-driven Talking Portrait</h3>
<div style="margin-top: 15px;">
<span style="margin-right: 15px; font-size: 1.3em;"><a href="https://taekyungki.github.io" target="_blank">Taekyung Ki<sup>1</sup></a></span>
<span style="margin-right: 15px; font-size: 1.3em;"><a href="https://kevinmin95.github.io" target="_blank">Dongchan Min<sup>2</sup></a></span>
<span style="margin-right: 15px; font-size: 1.3em;"><a href="https://www.aistudios.com/ko/home" target="_blank">Gyeongsu Chae<sup>1</sup></a></span>
</div>
<div style="margin-top: 15px;">
<span style="margin-right: 20px; font-size: 1.2em;"><sup>1</sup>DeepBrain AI Inc.</span>
<span style="margin-right: 20px; font-size: 1.2em;"><sup>2</sup>Graduate School of AI, KAIST</span>
</div>
<div>
<span style="margin-right: 10px; font-size: 1.3em;"> ArXiv 2024</span>
</div>
</div>
<div class="text-center" style="font-size: 1.5em; margin-top: 25px;">
<a class="btn btn-primary btn-lg" target="_blank"
href="https://arxiv.org/abs/2412.01064" role="button"
style="margin-right: 10px; margin-bottom: 10px;">Paper</a>
<a class="btn btn-primary btn-lg" target="_blank"
href="" role="button"
style="margin-right: 10px; margin-bottom: 10px;">Video (Soon)</a>
<!-- <a class="btn btn-primary btn-lg" target="_blank"
href="" role="button"
style="margin-right: 10px; margin-bottom: 10px;">Arxiv</a> -->
<a class="btn btn-primary btn-lg" target="_blank"
href="" role="button"
style="margin-bottom: 10px;">Code (Soon)</a>
</div>
<div style="margin-top: 30px;">
<h2 class="text-center">
Abstract
</h2>
<img src="./src/img/float-abstract.png" alt="" style="display: block; margin: 10px auto; width: 90%;">
<br>
<p style="font-style: italic; margin-bottom: 5px;">
With the rapid advancement of diffusion-based generative models, portrait image animation has achieved remarkable results.
However, it still faces challenges in temporally consistent video generation and fast sampling due to its iterative sampling nature.
This paper presents FLOAT, an audio-driven talking portrait video generation method based on flow matching generative model.
We shift the generative modeling from the pixel-based latent space to a learned motion latent space, enabling efficient design of temporally consistent motion.
To achieve this, we introduce a transformer-based vector field predictor with a simple yet effective frame-wise conditioning mechanism.
Additionally, our method supports speech-driven emotion enhancement, enabling a natural incorporation of expressive motions.
Extensive experiments demonstrate that our method outperforms state-of-the-art audio-driven talking portrait methods in terms of visual quality, motion fidelity, and efficiency.
</p>
<p style="font-size: 1.2em; margin-top: 0px;">
<span style="font-weight: bold;">TL;DR:</span> FLOAT is a flow matching based audio-driven talking portrait video generation method, which can enhance the speech-driven emotional motion.</h4>
</p>
</div>
<div style="margin-top: 50px;">
<div class="text-center">
<h2>
Method Overview
</h2>
<img src="./src/img/overview.png" width=100% class="img-fluid" alt="Responsive image">
</div>
<div style="margin-top: 35px;">
<p>
Audio-driven talking portrait aims to synthesize talking portrait videos using a single source portrait image and driving audio.
FLOAT is built upon a motion latent auto-encoder that encodes the given portrait image into an <span class="text-danger">identity-motion</span> latent representation.
We generates audio-conditioned talking portrait motion latents through the flow matching (with optimal transport trajectories).
To enhace the naturalness of generated talking motion, we incorporate the speech-driven emotion labels (😀), providing a natural approach of emotion-aware talking portrait motion generation.
</p>
</div>
</div>
<div style="margin-top:50px;">
<h2 class="text-center"> Results </h2>
<br>
<h4> Results with Out-of-distribution Datas</h4>
<p> FLOAT can generate realistic talking portrait videos using OOD portrait images and audio. </p>
<div class="video-container">
<video controls>
<source src="src/video/ood_musk.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<video controls>
<source src="src/video/ood_chinese.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<video controls>
<source src="src/video/ood_monarisa.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<br>
<div class="video-container">
<video controls>
<source src="src/video/ood_hinton.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<video controls>
<source src="src/video/ood_altman.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<video controls>
<source src="src/video/ood_paint.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<br>
<div class="video-container">
<video controls>
<source src="src/video/ood_feifei.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<video controls>
<source src="src/video/ood_taylor.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<video controls>
<source src="src/video/ood_paint_angy.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<br><br>
<h4>Emotion Redirection</h4>
<p> Since FLOAT is trained with speech-driven emotion labels, it can re-direct the emotion of the talking portrait during the inference phase.
Specifically, we can manipulate the predicted speech-driven emotion label with a simple one-hot emotion label, which can be further refined through classifier-free vector fields.
This enables users to refine emotion even when the driving speech conveys ambiguious or mixed emotions. </p>
<div class="row">
<div class="col-md-12 col-sm-12 col-xs-12 gallery">
<div style="position: relative; padding-top: 0; overflow: hidden;">
<video controls style="display: block; width: 100%; height: auto;">
<source src="src/video/emotion_redirection_1.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<br>
<div style="position: relative; padding-top: 0; overflow: hidden;">
<video controls style="display: block; width: 100%; height: auto;">
<source src="src/video/emotion_redirection_2.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
</div>
<br><br>
<h4>Comparison with State-of-the-art Methods</h4>
<p> We compare with state-of-the-art non-diffusion-based methods and diffusion-based methods.
For non-diffusion-based methods, we choose SadTalker and EDTalk.
For diffusion-based methods, we choose AniTalker, Hallo, and EchoMimic.</p>
<div id="carouselExampleControls" class="carousel slide" data-ride="carousel">
<div class="carousel-inner">
<div class="carousel-item active">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/sota_comparison_01.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline >
<source src="./src/video/sota_comparison_02.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/sota_comparison_03.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/sota_comparison_04.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/sota_comparison_05.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
<a class="carousel-control-prev" href="#carouselExampleControls" role="button" data-slide="prev">
<span class="carousel-control-prev-icon" aria-hidden="true"></span>
<span class="sr-only">Previous</span>
</a>
<a class="carousel-control-next" href="#carouselExampleControls" role="button" data-slide="next">
<span class="carousel-control-next-icon" aria-hidden="true"></span>
<span class="sr-only">Next</span>
</a>
</div>
<br>
<br>
<h4>Ablation Studies on Frame-wise AdaLN and Flow Matching</h4>
<p> We conduct ablation studies on frame-wise AdaLN (and gating) and flow matching. For frame-wise AdaLN (and gating), we compare it with cross-attention mechanism,
which is widely used in conditional generation. For flow matching, we compare it with two types of diffusion models (\(\epsilon\)-prediction and \(x_0\)-prediction).
We observe that frame-wise AdaLN (and gating) can generate more diverse head motions than cross-attention.
We also observe that flow mathcing can generate more temporally consistent videos with accurate lip-synchronization than diffusion models.
</p>
<div id="carouselExampleControls1" class="carousel slide" data-ride="carousel">
<div class="carousel-inner">
<div class="carousel-item active">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/ablation_phase2_architecture_1.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline >
<source src="./src/video/ablation_phase2_architecture_2.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/ablation_phase2_architecture_3.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/ablation_phase2_architecture_4.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
<a class="carousel-control-prev" href="#carouselExampleControls1" role="button" data-slide="prev">
<span class="carousel-control-prev-icon" aria-hidden="true"></span>
<span class="sr-only">Previous</span>
</a>
<a class="carousel-control-next" href="#carouselExampleControls1" role="button" data-slide="next">
<span class="carousel-control-next-icon" aria-hidden="true"></span>
<span class="sr-only">Next</span>
</a>
</div>
<br><br>
<h4> Different Number of Function Evaluations (NFEs)</h4>
<p> The small number of function evaluations (NFEs) affects temporal consistency. This is because we generate the motion latents, not the content itself.
FLOAT is capable of generating reasonable video results with approximately 10 NFEs. </p>
<div id="carouselExampleControls2" class="carousel slide" data-ride="carousel">
<div class="carousel-inner">
<div class="carousel-item active">
<video class="d-block w-100 video-fit" controls playsinline>
<source src="./src/video/ablation_phase2_NFE_1.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<div class="carousel-item">
<video class="d-block w-100 video-fit" controls playsinline >
<source src="./src/video/ablation_phase2_NFE_2.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
<a class="carousel-control-prev" href="#carouselExampleControls2" role="button" data-slide="prev">
<span class="carousel-control-prev-icon" aria-hidden="true"></span>
<span class="sr-only">Previous</span>
</a>
<a class="carousel-control-next" href="#carouselExampleControls2" role="button" data-slide="next">
<span class="carousel-control-next-icon" aria-hidden="true"></span>
<span class="sr-only">Next</span>
</a>
</div>
<br>
<br>
<h4>Emotion Guidance Scales</h4>
<p> We can control the intensity of the emotion by adjusting the emotion guidance scale.
Note that the predicted speech-driven emotion label is <b>Disgust</b> with a 99.99% probability. </p>
<div class="row">
<div class="col-md-12 col-sm-12 col-xs-12 gallery">
<div style="position: relative; padding-top: 0; overflow: hidden;">
<video controls style="display: block; width: 100%; height: auto;">
<source src="src/video/ablation_phase2_emotion-guidance.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
</div>
<br>
<br>
<h4>Additional Driving Conditions</h4>
<p> We also investigate another types of driving conditions, such as 3DMM head pose parameters, to improve the controllability and naturalness.
Here, 3DPose, S2E, and I2E are 3DMM head pose parameters, Speech-to-emotion label, and Image-to-emotion label, resepctively.
</p>
<div class="row">
<div class="col-md-12 col-sm-12 col-xs-12 gallery">
<div style="position: relative; padding-top: 0; overflow: hidden;">
<video controls style="display: block; width: 100%; height: auto;">
<source src="src/video/driving_condition_01.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
<br>
<div style="position: relative; padding-top: 0; overflow: hidden;">
<video controls style="display: block; width: 100%; height: auto;">
<source src="src/video/driving_condition_02.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
</div>
<br>
<br>
<h4>Ablation Study on Facial Component Perceptual Loss in Phase 1</h4>
<p> The proposed facial component perceptual loss for the motion latent auto-encoder
significantly improves visual quality (e.g., teeth and eyes), as well as fine-grained motion (e.g., eyeball movement and lip motion).</p>
<div class="row">
<div class="col-md-12 col-sm-12 col-xs-12 gallery">
<div style="position: relative; padding-top: 0; overflow: hidden;">
<video controls muted style="display: block; width: 100%; height: auto;">
<source src="src/video/ablation_phase1_facial_component_perceptual_loss.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
</div>
</div>
</div>
</div>
<br>
<div>
<h2 class="text-center" style="margin-top: 30px;">
Citation
</h2>
<p>
If you want to cite our work, please use:
</p>
<pre>
@article{ki2024float,
title={FLOAT: Generative Motion Latent Flow Matching for Audio-driven Talking Portrait},
author={Ki, Taekyung and Min, Dongchan and Chae, Gyeongsu},
journal={arXiv preprint arXiv:2412.01064},
year={2024}
}
</pre>
</div>
<div>
<h2 class="text-center" style="margin-top: 30px;">
Acknowledgement
</h2>
<p>
The source images and audio are collected from the internet and other baselines, such as SadTalker, EMO, VASA-1, Hallo, LivePortrait, Loopy, and others.
We appreciate their valuable contributions to this field.
This project page is based on the project page of <a href="https://m-niemeyer.github.io/regnerf">RegNeRF</a>. You can easily use it from <a href="https://github.com/m-niemeyer/regnerf" target="_blank">the github repository</a>.
</p>
</div>
<br>
<br>
<br>
<script src="https://code.jquery.com/jquery-3.2.1.slim.min.js"
integrity="sha384-KJ3o2DKtIkvYIK3UENzmM7KCkRr/rE9/Qpg6aAZGJwFDMVNA/GpGFF93hXpG5KkN"
crossorigin="anonymous"></script>
<script src="https://cdnjs.cloudflare.com/ajax/libs/popper.js/1.12.9/umd/popper.min.js"
integrity="sha384-ApNbgh9B+Y1QKtv3Rn7W3mgPxhU9K/ScQsAP7hUibX39j7fakFPskvXusvfa0b4Q"
crossorigin="anonymous"></script>
<script src="https://maxcdn.bootstrapcdn.com/bootstrap/4.0.0/js/bootstrap.min.js"
integrity="sha384-JZR6Spejh4U02d8jOt6vLEHfe/JQGiRRSQQxSfFWpi1MquVdAyjUar5+76PVCmYl"
crossorigin="anonymous"></script>
<script>
const carousel = document.querySelector('#carouselExampleControls');
const videos = carousel.querySelectorAll('video');
videos.forEach(video => {
video.addEventListener('play', () => {
$(carousel).carousel('pause');
});
video.addEventListener('pause', () => {
$(carousel).carousel('cycle');
});
video.addEventListener('ended', () => {
$(carousel).carousel('cycle');
});
});
</script>
<script>
const carousels = document.querySelectorAll('.carousel');
carousels.forEach(carousel => {
const videos = carousel.querySelectorAll('video');
videos.forEach(video => {
video.addEventListener('play', () => {
$(carousel).carousel('pause');
});
video.addEventListener('pause', () => {
$(carousel).carousel('cycle');
});
video.addEventListener('ended', () => {
$(carousel).carousel('cycle');
});
});
});
</script>
<script type="text/javascript" async
src="https://cdnjs.cloudflare.com/ajax/libs/mathjax/2.7.7/MathJax.js?config=TeX-MML-AM_CHTML">
</script>
</body>
</html>