-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
836 lines (796 loc) · 43.9 KB
/
Copy pathindex.html
File metadata and controls
836 lines (796 loc) · 43.9 KB
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
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8" />
<meta name="viewport" content="width=device-width, initial-scale=1" />
<title>OmniVAE: An Audio-Video VAE with Cross-Modal Alignment for Joint Generation</title>
<meta
name="description"
content="OmniVAE jointly trains audio and video VAEs with segment-level contrastive alignment and modality-specific semantic distillation for joint generation."
/>
<meta property="og:title" content="OmniVAE" />
<meta
property="og:description"
content="An Audio-Video VAE with Cross-Modal Alignment for Joint Generation."
/>
<meta property="og:type" content="website" />
<meta property="og:image" content="assets/images/omnivae-architecture.png" />
<link rel="stylesheet" href="styles.css?v=18" />
</head>
<body>
<header class="site-header" aria-label="Primary navigation">
<div class="nav-container">
<a class="brand" href="#top">OmniVAE</a>
<nav>
<a href="#abstract">Abstract</a>
<a href="#method">Method</a>
<a href="#results">Results</a>
<a href="#cases">Cases</a>
<a href="#citation">Citation</a>
</nav>
</div>
</header>
<main id="top">
<section class="hero" aria-labelledby="hero-title">
<div class="container hero-content">
<p class="hero-kicker">Research project · Audio-video representation learning</p>
<h1 id="hero-title">
<span>OmniVAE</span>: An Audio-Video VAE<br class="title-break" />
with Cross-Modal Alignment<br class="title-break" />
for Joint Generation
</h1>
<div class="authors" aria-label="Authors">
<span><a href="https://junzhan2000.github.io/">Jun Zhan</a><sup>1,2,3,*,‡</sup></span>
<span>Chen Yang<sup>1,2,3,*</sup></span>
<span>Yitian Gong<sup>1,3,*</sup></span>
<span>Donghua Yu<sup>1,3</sup></span>
<span>Kuangwei Chen<sup>1,3</sup></span>
<span>Wenbo Zhang<sup>1,3</sup></span>
<span>Kexin Huang<sup>1,3</sup></span>
<span>Qi Luo<sup>1,3</sup></span>
<span>Zhe Xu<sup>1,3</sup></span>
<span>Ying Zhu<sup>3,4</sup></span>
<span>Jin Wang<sup>1,3</sup></span>
<span>Tengyue Zhang<sup>2,4</sup></span>
<span>Qi Chen<sup>2,4</sup></span>
<span>Cheng Chang<sup>1,3</sup></span>
<span>Songlin Wang<sup>3</sup></span>
<span>Junqi Dai<sup>1</sup></span>
<span>Jiasheng Ye<sup>1</sup></span>
<span>Xiaogui Yang<sup>3</sup></span>
<span>Tianyi Liang<sup>2,3</sup></span>
<span>Xiangyu Peng<sup>3</sup></span>
<span>Zhaoye Fei<sup>1,2,3</sup></span>
<span>Shimin Li<sup>1,3</sup></span>
<span>Qinyuan Cheng<sup>1,3</sup></span>
<span>Xie Chen<sup>2,4</sup></span>
<span>Xinchi Chen<sup>1</sup></span>
<span>Xipeng Qiu<sup>1,2,3,†</sup></span>
</div>
<div class="affiliations">
<span><sup>1</sup>Fudan University</span>
<span><sup>2</sup>Shanghai Innovation Institute</span>
<span><sup>3</sup>MOSI Intelligence</span>
<span><sup>4</sup>Shanghai Jiao Tong University</span>
</div>
<p class="author-notes">
<sup>*</sup>Equal contribution <sup>‡</sup>Project lead
<sup>†</sup>Corresponding author
</p>
<div class="actions" aria-label="Project links">
<a class="button" href="assets/paper/OmniVAE.pdf">
<svg aria-hidden="true" viewBox="0 0 24 24"><path d="M6 2h8l4 4v16H6V2zm7 1.4V7h3.6L13 3.4zM8.5 11h7v1.5h-7V11zm0 3.2h7v1.5h-7v-1.5zm0-6.4h3.5v1.5H8.5V7.8z"/></svg>
<span>Paper</span>
</a>
<a class="button" href="https://github.com/OpenMOSS/OmniVAE">
<svg aria-hidden="true" viewBox="0 0 24 24"><path d="M12 2a10 10 0 0 0-3.16 19.49c.5.09.68-.22.68-.48v-1.87c-2.78.6-3.37-1.18-3.37-1.18-.45-1.16-1.11-1.47-1.11-1.47-.91-.62.07-.61.07-.61 1 .07 1.53 1.03 1.53 1.03.9 1.53 2.34 1.09 2.91.83.09-.65.35-1.09.64-1.34-2.22-.25-4.55-1.11-4.55-4.94 0-1.09.39-1.98 1.03-2.68-.1-.25-.45-1.27.1-2.64 0 0 .84-.27 2.75 1.02A9.57 9.57 0 0 1 12 6.82c.85 0 1.71.11 2.51.34 1.91-1.29 2.75-1.02 2.75-1.02.55 1.37.2 2.39.1 2.64.64.7 1.03 1.59 1.03 2.68 0 3.84-2.34 4.69-4.57 4.94.36.31.68.92.68 1.85v2.76c0 .27.18.58.69.48A10 10 0 0 0 12 2Z"/></svg>
<span>Code</span>
</a>
<a class="button" href="https://huggingface.co/OpenMOSS-Team/OmniVAE">
<img class="hf-icon" src="assets/images/huggingface-logo.svg" alt="" />
<span>Models</span>
</a>
</div>
</div>
</section>
<section id="abstract" class="section">
<div class="container reading-width">
<div class="section-heading centered">
<p class="section-label">Abstract</p>
<h2>Aligning audio and video in VAE latents</h2>
</div>
<div class="abstract-text">
<p>
Recent generative models are moving beyond silent video or
standalone audio synthesis toward the joint generation of
synchronized audio and video. Most existing systems use audio and
video VAEs trained separately, leaving the downstream generator to
learn cross-modal synchronization from scratch.
</p>
<p>
OmniVAE is a jointly trained audio-video VAE that learns
fine-grained semantic alignment between audio and video latent
representations. A segment-level audio-video contrastive objective
captures temporal-semantic correspondence, while semantic features
from pretrained modality-specific encoders are distilled into each
branch to improve downstream learnability. Both objectives are
training-only and add no inference cost. Experiments show improved
latent-space probing, generation quality, and synchronization in
downstream text-to-audio-video generation.
</p>
</div>
</div>
</section>
<section id="method" class="section band">
<div class="container">
<div class="section-heading centered">
<p class="section-label">Method</p>
<h2>Separate latents, explicitly aligned latent spaces</h2>
<p>
OmniVAE preserves modality-specific reconstruction while making
audio and video latents easier to model individually and jointly.
</p>
</div>
<figure class="paper-figure architecture">
<a href="assets/images/omnivae-architecture.png">
<img
src="assets/images/omnivae-architecture.png"
alt="OmniVAE architecture with independent video and audio VAEs, audio-visual contrastive learning, and uni-modality semantic distillation."
/>
</a>
<figcaption>
<strong>Overview of OmniVAE.</strong> Separate video and audio VAEs
retain their standard reconstruction paths. Segment-level
audio-visual contrastive learning and uni-modality semantic
distillation supervise the latents only during training.
</figcaption>
</figure>
</div>
</section>
<section id="results" class="section results-section">
<div class="container">
<div class="section-heading centered">
<p class="section-label">Results</p>
<h2>Cross-modal alignment improves downstream joint generation</h2>
<p>
Results below reproduce the paper's T2AV generation, reconstruction,
and audio-video sync probing evaluations.
</p>
</div>
<section class="result-block featured-result" aria-labelledby="t2av-title">
<div class="result-heading">
<p class="result-index">01 · Text-to-Audio-Video Generation</p>
<h3 id="t2av-title">Verse-Bench T2AV evaluation</h3>
<p>
Models trained with AVCLIP consistently improve temporal and
semantic alignment. Combining AVCLIP with semantic distillation
gives the strongest overall audio performance and cross-modal
alignment.
</p>
</div>
<figure class="paper-figure curve-figure">
<a href="assets/images/all.pdf">
<img
src="assets/images/t2av-training-curves.png"
alt="Sixteen T2AV evaluation curves comparing Recon, Recon plus Distill, Recon plus AVCLIP, and OmniVAE across training steps."
/>
</a>
<figcaption>
<strong>Figure 2.</strong> T2AV metrics across training steps.
Rows show cross-modal alignment, video quality, audio quality,
and AudioBox aesthetics. Click the figure to open the original
<code>all.pdf</code>.
</figcaption>
</figure>
<div class="table-grid t2av-tables">
<article class="table-panel">
<h4>Cross-Modal Alignment</h4>
<div class="table-scroll">
<table>
<thead>
<tr><th>Configuration</th><th>DeSync ↓</th><th>LSE-C ↑</th><th>V-A Sim ↑</th><th>T-A Sim ↑</th></tr>
</thead>
<tbody>
<tr><td>Recon</td><td>0.884</td><td>1.479</td><td>0.254</td><td>0.201</td></tr>
<tr><td>Recon + Distill</td><td>0.814</td><td>1.450</td><td>0.256</td><td>0.227</td></tr>
<tr><td>Recon + AVCLIP</td><td>0.576</td><td>1.970</td><td>0.257</td><td>0.225</td></tr>
<tr class="best"><td>OmniVAE</td><td>0.570</td><td>2.093</td><td>0.274</td><td>0.246</td></tr>
</tbody>
</table>
</div>
</article>
<article class="table-panel">
<h4>Video Quality</h4>
<div class="table-scroll">
<table>
<thead>
<tr><th>Configuration</th><th>Aesthetic ↑</th><th>MusiQ ↑</th><th>ManiQA ↑</th><th>Motion ↑</th></tr>
</thead>
<tbody>
<tr><td>Recon</td><td>0.369</td><td>0.503</td><td>0.338</td><td class="metric-best">0.556</td></tr>
<tr><td>Recon + Distill</td><td>0.373</td><td class="metric-best">0.523</td><td class="metric-best">0.359</td><td>0.495</td></tr>
<tr><td>Recon + AVCLIP</td><td class="metric-best">0.382</td><td>0.503</td><td>0.341</td><td>0.533</td></tr>
<tr class="ours"><td>OmniVAE</td><td>0.381</td><td>0.515</td><td>0.355</td><td>0.456</td></tr>
</tbody>
</table>
</div>
</article>
<article class="table-panel">
<h4>Audio Quality</h4>
<div class="table-scroll">
<table>
<thead>
<tr><th>Configuration</th><th>IS ↑</th><th>KL ↓</th><th>FD ↓</th><th>WER ↓</th></tr>
</thead>
<tbody>
<tr><td>Recon</td><td>3.041</td><td>1.279</td><td>1.029</td><td>0.243</td></tr>
<tr><td>Recon + Distill</td><td>3.604</td><td>1.258</td><td>0.959</td><td>0.205</td></tr>
<tr><td>Recon + AVCLIP</td><td>3.598</td><td>1.231</td><td>0.932</td><td>0.172</td></tr>
<tr class="best"><td>OmniVAE</td><td>4.001</td><td>1.140</td><td>0.875</td><td>0.168</td></tr>
</tbody>
</table>
</div>
</article>
<article class="table-panel">
<h4>AudioBox Aesthetics</h4>
<div class="table-scroll">
<table>
<thead>
<tr><th>Configuration</th><th>CE ↑</th><th>CU ↑</th><th>PC ↓</th><th>PQ ↑</th></tr>
</thead>
<tbody>
<tr><td>Recon</td><td>4.625</td><td>5.855</td><td>2.282</td><td>6.328</td></tr>
<tr><td>Recon + Distill</td><td>4.533</td><td>5.677</td><td class="metric-best">2.232</td><td>6.060</td></tr>
<tr><td>Recon + AVCLIP</td><td>4.666</td><td>5.894</td><td>2.237</td><td>6.219</td></tr>
<tr class="ours"><td>OmniVAE</td><td class="metric-best">4.826</td><td class="metric-best">6.119</td><td>2.259</td><td class="metric-best">6.406</td></tr>
</tbody>
</table>
</div>
</article>
</div>
<p class="table-note">
<strong>Table 4.</strong> Mean over checkpoints 180k, 190k, and
200k with CFG = 5. LSE-C is evaluated on face-detectable Set3
samples. CE, CU, PC, and PQ denote Content Enjoyment, Content
Usefulness, Production Complexity, and Production Quality.
</p>
</section>
<section class="result-block" aria-labelledby="recon-title">
<div class="result-heading">
<p class="result-index">02 · Reconstruction</p>
<h3 id="recon-title">Reconstruction quality is largely preserved</h3>
<p>
Alignment and semantic objectives keep video reconstruction close
to the reconstruction-only baseline. OmniVAE also remains
competitive across speech, general audio, and music benchmarks.
</p>
</div>
<article class="table-panel full-width">
<h4>Video Reconstruction · UCF-101 / Panda-70M</h4>
<div class="table-scroll">
<table>
<thead>
<tr><th>Configuration</th><th>PSNR ↑</th><th>SSIM ↑</th><th>LPIPS ↓</th><th>L1 ↓</th><th>rFVD ↓</th></tr>
</thead>
<tbody>
<tr class="group-row"><td colspan="6">External VAEs</td></tr>
<tr><td>Wan2.1 VAE</td><td>34.50 / 32.62</td><td>0.9510 / 0.9448</td><td>0.0201 / 0.0169</td><td>0.0117 / 0.0130</td><td>2.46 / 1.49</td></tr>
<tr><td>Wan2.2 VAE</td><td>34.87 / 33.21</td><td>0.9584 / 0.9541</td><td>0.0177 / 0.0137</td><td>0.0110 / 0.0120</td><td>3.72 / 1.38</td></tr>
<tr class="group-row"><td colspan="6">Ours</td></tr>
<tr><td>Recon</td><td>36.93 / 36.56</td><td>0.9656 / 0.9697</td><td>0.0102 / 0.0067</td><td>0.0090 / 0.0086</td><td>2.40 / 1.25</td></tr>
<tr><td>Recon + Distill</td><td>36.70 / 36.22</td><td>0.9649 / 0.9688</td><td>0.0107 / 0.0071</td><td>0.0092 / 0.0089</td><td>2.53 / 1.26</td></tr>
<tr><td>Recon + AVCLIP</td><td>36.22 / 35.13</td><td>0.9616 / 0.9631</td><td>0.0115 / 0.0079</td><td>0.0097 / 0.0098</td><td>2.66 / 1.35</td></tr>
<tr class="ours"><td>OmniVAE</td><td>36.27 / 35.24</td><td>0.9616 / 0.9632</td><td>0.0113 / 0.0077</td><td>0.0097 / 0.0098</td><td>2.81 / 1.31</td></tr>
</tbody>
</table>
</div>
</article>
<article class="table-panel full-width">
<h4>Audio Reconstruction · Speech and Audio / Music</h4>
<div class="table-scroll">
<table>
<thead>
<tr><th>Model</th><th>SIM ↑</th><th>STOI ↑</th><th>P-NB ↑</th><th>P-WB ↑</th><th>Mel ↓</th><th>STFT ↓</th><th>ViSQOL ↑</th></tr>
</thead>
<tbody>
<tr class="group-row"><td colspan="8">External VAEs</td></tr>
<tr><td>HunyuanVideo-Foley VAE</td><td>0.9899</td><td>0.9979</td><td>4.4780</td><td>4.5337</td><td>0.52 / 0.60</td><td>1.82 / 1.84</td><td>4.47 / 4.45</td></tr>
<tr><td>MMAudio VAE</td><td>0.9122</td><td>0.9502</td><td>3.4931</td><td>2.9797</td><td>1.23 / 1.36</td><td>2.14 / 2.44</td><td>4.42 / 4.42</td></tr>
<tr><td>Stable Audio 3 SAME-S</td><td>0.7960</td><td>0.9139</td><td>2.8073</td><td>2.2991</td><td>1.46 / 1.37</td><td>3.19 / 3.67</td><td>3.68 / 3.99</td></tr>
<tr><td>Stable Audio 3 SAME-L</td><td>0.8755</td><td>0.9539</td><td>3.4615</td><td>3.0739</td><td>1.37 / 1.22</td><td>3.30 / 3.55</td><td>3.76 / 4.17</td></tr>
<tr class="group-row"><td colspan="8">Ours</td></tr>
<tr><td>Recon</td><td>0.9893</td><td>0.9981</td><td>4.4916</td><td>4.5557</td><td>0.44 / 0.50</td><td>1.75 / 1.72</td><td>4.55 / 4.60</td></tr>
<tr><td>Recon + Distill</td><td>0.9893</td><td>0.9970</td><td>4.4453</td><td>4.5101</td><td>0.56 / 0.64</td><td>1.96 / 1.95</td><td>4.44 / 4.42</td></tr>
<tr><td>Recon + AVCLIP</td><td>0.9666</td><td>0.9946</td><td>4.4346</td><td>4.3638</td><td>0.76 / 0.75</td><td>1.99 / 2.03</td><td>4.03 / 4.18</td></tr>
<tr class="ours"><td>OmniVAE</td><td>0.9737</td><td>0.9948</td><td>4.4400</td><td>4.4085</td><td>0.74 / 0.73</td><td>2.02 / 2.08</td><td>4.08 / 4.31</td></tr>
</tbody>
</table>
</div>
</article>
<p class="table-note">
<strong>Table 2.</strong> Video values are UCF-101 / Panda-70M;
Audio / Music values are AudioSet / MUSDB18. Speech metrics use
LibriSpeech test-clean.
</p>
</section>
<section class="result-block probe-result" aria-labelledby="probe-title">
<div class="result-heading">
<p class="result-index">03 · Audio-Video Sync Probing</p>
<h3 id="probe-title">Contrastive learning drives temporal alignment</h3>
<p>
A lightweight probe predicts one of 21 audio offsets from −2 to
+2 seconds on VGGSound-Sparse. OmniVAE performs best with frozen
representations and when the contrastive aggregators are
fine-tuned.
</p>
</div>
<article class="table-panel probe-table">
<div class="table-scroll">
<table>
<thead>
<tr><th>Protocol</th><th>Method</th><th>A@1</th><th>A@5</th><th>A@1<sub>tol</sub></th></tr>
</thead>
<tbody>
<tr><td rowspan="4">Frozen</td><td>Recon</td><td>6.4</td><td>31.5</td><td>16.2</td></tr>
<tr><td>Recon + Distill</td><td>7.1</td><td>33.0</td><td>17.5</td></tr>
<tr><td>Recon + AVCLIP</td><td>18.1</td><td>47.8</td><td>34.5</td></tr>
<tr class="best"><td>OmniVAE</td><td>20.2</td><td>50.1</td><td>37.2</td></tr>
<tr><td rowspan="4">+ft</td><td>Recon + ft</td><td>7.0</td><td>32.4</td><td>17.1</td></tr>
<tr><td>Recon + Distill + ft</td><td>7.8</td><td>33.9</td><td>18.4</td></tr>
<tr><td>Recon + AVCLIP + ft</td><td>21.8</td><td>52.4</td><td>39.6</td></tr>
<tr class="best"><td>OmniVAE + ft</td><td>22.3</td><td>53.2</td><td>40.4</td></tr>
</tbody>
</table>
</div>
</article>
<p class="table-note">
<strong>Table 3.</strong> Sync probing on VGGSound-Sparse (%).
A@1<sub>tol</sub> uses ±1-class tolerance; <em>+ft</em> also tunes
the contrastive aggregators.
</p>
</section>
</div>
</section>
<section id="cases" class="section case-studies">
<div class="container">
<div class="section-heading">
<p class="section-label">Qualitative Comparisons</p>
<h2>Qualitative comparisons across Verse-Bench settings</h2>
<p>
Each case compares the same prompt and sampling setting across four
training objectives at 200k steps (CFG = 5). Core per-sample metrics
are shown first; expand a case to inspect the complete evaluation.
Set3 labels refer to camera framing, not dataset scale.
</p>
</div>
<div class="case-list">
<div class="case-group">
<h3>Set1</h3>
<p>Non-talking or non-speech-leaning samples with strong synchronization gains.</p>
</div>
<article class="case-card" data-case="set1-0070">
<div class="case-header">
<span>Set1 / 0070</span>
<h3>A calm home-office presentation delivered in front of a bookshelf.</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set1-0070-recon.jpg">
<source src="assets/videos/cases/set1-0070-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set1-0070-omnivae.jpg">
<source src="assets/videos/cases/set1-0070-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.80</td><td>0.20</td><td>1.60 lower</td></tr>
<tr><td>Video AS ↑</td><td>0.419</td><td>0.427</td><td>+0.008</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.382</td><td>0.410</td><td>+0.029</td></tr>
<tr><td>DNSMOS ↑</td><td>2.97</td><td>3.97</td><td>+0.99</td></tr>
<tr><td>Audio PQ ↑</td><td>4.08</td><td>5.83</td><td>+1.74</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set1-0109">
<div class="case-header">
<span>Set1 / 0109 · music</span>
<h3>Intimate piano performance in a small studio.</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set1-0109-recon.jpg">
<source src="assets/videos/cases/set1-0109-recon.mp4" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set1-0109-omnivae.jpg">
<source src="assets/videos/cases/set1-0109-omnivae.mp4" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.80</td><td>0.00</td><td>1.80 lower</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.001</td><td>0.124</td><td>+0.123</td></tr>
<tr><td>DNSMOS ↑</td><td>2.33</td><td>2.63</td><td>+0.30</td></tr>
</tbody>
</table>
</div>
</article>
<div class="case-group">
<h3>Set2</h3>
<p>Longer spoken prompts where aligned latents improve synchronization and audio fidelity.</p>
</div>
<article class="case-card" data-case="set2-10000344">
<div class="case-header">
<span>Set2 / 10000344</span>
<h3>"Long time ago, every year when I watch Oscar with my..."</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set2-10000344-recon.jpg">
<source src="assets/videos/cases/set2-10000344-recon.mp4" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set2-10000344-omnivae.jpg">
<source src="assets/videos/cases/set2-10000344-omnivae.mp4" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.90</td><td>0.10</td><td>1.80 lower</td></tr>
<tr><td>Video AS ↑</td><td>0.493</td><td>0.517</td><td>+0.024</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.263</td><td>0.294</td><td>+0.032</td></tr>
<tr><td>DNSMOS ↑</td><td>3.83</td><td>4.00</td><td>+0.18</td></tr>
<tr><td>Audio FD ↓</td><td>0.749</td><td>0.498</td><td>0.252 lower</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set2-10000292">
<div class="case-header">
<span>Set2 / 10000292</span>
<h3>"It was great. I have great relationships with Mexico, with the Mexican people."</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set2-10000292-recon.jpg">
<source src="assets/videos/cases/set2-10000292-recon.mp4" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set2-10000292-omnivae.jpg">
<source src="assets/videos/cases/set2-10000292-omnivae.mp4" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.40</td><td>0.10</td><td>1.30 lower</td></tr>
<tr><td>Video AS ↑</td><td>0.341</td><td>0.364</td><td>+0.023</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.308</td><td>0.344</td><td>+0.036</td></tr>
<tr><td>DNSMOS ↑</td><td>3.71</td><td>3.85</td><td>+0.14</td></tr>
<tr><td>Speech WER ↓</td><td>0.077</td><td>0.000</td><td>0.077 lower</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set2-10000519">
<div class="case-header">
<span>Set2 / 10000519</span>
<h3>Formal conference presentation with a single speaker.</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set2-10000519-recon.jpg">
<source src="assets/videos/cases/set2-10000519-recon.mp4" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set2-10000519-omnivae.jpg">
<source src="assets/videos/cases/set2-10000519-omnivae.mp4" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.50</td><td>0.10</td><td>1.40 lower</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.359</td><td>0.367</td><td>+0.007</td></tr>
<tr><td>DNSMOS ↑</td><td>3.12</td><td>3.59</td><td>+0.47</td></tr>
<tr><td>Audio FD ↓</td><td>1.132</td><td>1.013</td><td>0.120 lower</td></tr>
<tr><td>Audio PQ ↑</td><td>5.81</td><td>6.72</td><td>+0.91</td></tr>
</tbody>
</table>
</div>
</article>
<div class="case-group">
<h3>Set3 · Medium Shot</h3>
<p>Medium close-up framing keeps the speaker’s upper body and some stage context visible while enlarging the face for lip-sync evaluation.</p>
</div>
<article class="case-card" data-case="set3m-1006">
<div class="case-header">
<span>Set3 · Medium Shot / 1006</span>
<h3>"Wouldn't it be great if there was a way to meet that demand by simply wasting less of what we're already mining?"</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set3m-1006-recon.jpg">
<source src="assets/videos/cases/set3m-1006-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set3m-1006-omnivae.jpg">
<source src="assets/videos/cases/set3m-1006-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.90</td><td>0.20</td><td>1.70 lower</td></tr>
<tr><td>Video AS ↑</td><td>0.399</td><td>0.412</td><td>+0.013</td></tr>
<tr><td>PE-TAV ↑</td><td>3.46</td><td>3.55</td><td>+0.10</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set3m-1032">
<div class="case-header">
<span>Set3 · Medium Shot / 1032</span>
<h3>"So in order to understand how to actually achieve this, we have to look at cement in a little bit more detail."</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set3m-1032-recon.jpg">
<source src="assets/videos/cases/set3m-1032-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set3m-1032-omnivae.jpg">
<source src="assets/videos/cases/set3m-1032-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.20</td><td>0.00</td><td>1.20 lower</td></tr>
<tr><td>LSE-C ↑</td><td>1.19</td><td>1.88</td><td>+0.69</td></tr>
<tr><td>Video AS ↑</td><td>0.379</td><td>0.437</td><td>+0.058</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set3m-1045">
<div class="case-header">
<span>Set3 · Medium Shot / 1045</span>
<h3>"But in reality, I started it because I was traumatized, and I didn't want anyone to go through what I had been through."</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set3m-1045-recon.jpg">
<source src="assets/videos/cases/set3m-1045-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set3m-1045-omnivae.jpg">
<source src="assets/videos/cases/set3m-1045-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.10</td><td>0.00</td><td>1.10 lower</td></tr>
<tr><td>LSE-C ↑</td><td>1.30</td><td>1.54</td><td>+0.24</td></tr>
<tr><td>Video AS ↑</td><td>0.424</td><td>0.430</td><td>+0.006</td></tr>
<tr><td>DNSMOS ↑</td><td>3.22</td><td>3.30</td><td>+0.08</td></tr>
<tr><td>Audio PQ ↑</td><td>5.11</td><td>5.42</td><td>+0.32</td></tr>
</tbody>
</table>
</div>
</article>
<div class="case-group">
<h3>Set3 · Large Shot</h3>
<p>Close-up framing keeps the speaker’s face dominant in the frame, improving face detectability for lip-sync evaluation.</p>
</div>
<article class="case-card" data-case="set3l-1017">
<div class="case-header">
<span>Set3 · Large Shot / 1017</span>
<h3>A tightly framed speaker delivers a talk against a softly blurred blue stage backdrop.</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set3l-1017-recon.jpg">
<source src="assets/videos/cases/set3l-1017-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set3l-1017-omnivae.jpg">
<source src="assets/videos/cases/set3l-1017-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.90</td><td>0.30</td><td>1.60 lower</td></tr>
<tr><td>LSE-C ↑</td><td>2.22</td><td>3.18</td><td>+0.96</td></tr>
<tr><td>Video AS ↑</td><td>0.366</td><td>0.388</td><td>+0.022</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.201</td><td>0.338</td><td>+0.138</td></tr>
<tr><td>Speech WER ↓</td><td>0.200</td><td>0.133</td><td>0.067 lower</td></tr>
<tr><td>Audio PQ ↑</td><td>4.67</td><td>5.99</td><td>+1.32</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set3l-1041">
<div class="case-header">
<span>Set3 · Large Shot / 1041</span>
<h3>“When Americans see that suddenly prices of goods have gone way up.”</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set3l-1041-recon.jpg">
<source src="assets/videos/cases/set3l-1041-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set3l-1041-omnivae.jpg">
<source src="assets/videos/cases/set3l-1041-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>1.80</td><td>0.20</td><td>1.60 lower</td></tr>
<tr><td>LSE-C ↑</td><td>1.06</td><td>2.09</td><td>+1.03</td></tr>
<tr><td>Video AS ↑</td><td>0.394</td><td>0.426</td><td>+0.032</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.185</td><td>0.259</td><td>+0.074</td></tr>
<tr><td>DNSMOS ↑</td><td>3.34</td><td>3.74</td><td>+0.39</td></tr>
<tr><td>Audio PQ ↑</td><td>6.78</td><td>7.11</td><td>+0.33</td></tr>
</tbody>
</table>
</div>
</article>
<article class="case-card" data-case="set3l-1002">
<div class="case-header">
<span>Set3 · Large Shot / 1002</span>
<h3>“The get married advocates like to point to data that show that married people report higher life satisfaction.”</h3>
</div>
<div class="case-body">
<div class="case-videos">
<figure>
<video controls preload="metadata" poster="assets/videos/cases/set3l-1002-recon.jpg">
<source src="assets/videos/cases/set3l-1002-recon.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon</figcaption>
</figure>
<figure class="case-best">
<video controls preload="metadata" poster="assets/videos/cases/set3l-1002-omnivae.jpg">
<source src="assets/videos/cases/set3l-1002-omnivae.mp4?v=2" type="video/mp4" />
</video>
<figcaption>Recon + Distill + AVCLIP</figcaption>
</figure>
</div>
<table class="case-table">
<thead>
<tr><th>Metric</th><th>Recon</th><th>OmniVAE</th><th>Change</th></tr>
</thead>
<tbody>
<tr><td>DeSync ↓</td><td>2.00</td><td>0.20</td><td>1.80 lower</td></tr>
<tr><td>LSE-C ↑</td><td>0.89</td><td>1.95</td><td>+1.06</td></tr>
<tr><td>Video AS ↑</td><td>0.344</td><td>0.426</td><td>+0.082</td></tr>
<tr><td>Audio-text CLAP ↑</td><td>0.156</td><td>0.198</td><td>+0.042</td></tr>
<tr><td>PE-TAV ↑</td><td>3.10</td><td>3.64</td><td>+0.54</td></tr>
<tr><td>Audio PQ ↑</td><td>5.65</td><td>5.81</td><td>+0.16</td></tr>
</tbody>
</table>
</div>
</article>
</div>
</div>
</section>
<section id="resources" class="section band">
<div class="container resource-layout">
<div class="resource-intro">
<p class="section-label">Resources</p>
<p>Paper, code, and pretrained models</p>
</div>
<div class="resource-list">
<a href="assets/paper/OmniVAE.pdf">
<svg aria-hidden="true" viewBox="0 0 24 24"><path d="M6 2h8l4 4v16H6V2zm7 1.4V7h3.6L13 3.4zM8.5 11h7v1.5h-7V11zm0 3.2h7v1.5h-7v-1.5zm0-6.4h3.5v1.5H8.5V7.8z"/></svg>
<strong>Paper</strong>
</a>
<a href="https://github.com/OpenMOSS/OmniVAE">
<svg aria-hidden="true" viewBox="0 0 24 24"><path d="M12 2a10 10 0 0 0-3.16 19.49c.5.09.68-.22.68-.48v-1.87c-2.78.6-3.37-1.18-3.37-1.18-.45-1.16-1.11-1.47-1.11-1.47-.91-.62.07-.61.07-.61 1 .07 1.53 1.03 1.53 1.03.9 1.53 2.34 1.09 2.91.83.09-.65.35-1.09.64-1.34-2.22-.25-4.55-1.11-4.55-4.94 0-1.09.39-1.98 1.03-2.68-.1-.25-.45-1.27.1-2.64 0 0 .84-.27 2.75 1.02A9.57 9.57 0 0 1 12 6.82c.85 0 1.71.11 2.51.34 1.91-1.29 2.75-1.02 2.75-1.02.55 1.37.2 2.39.1 2.64.64.7 1.03 1.59 1.03 2.68 0 3.84-2.34 4.69-4.57 4.94.36.31.68.92.68 1.85v2.76c0 .27.18.58.69.48A10 10 0 0 0 12 2Z"/></svg>
<strong>Code</strong>
</a>
<a href="https://huggingface.co/OpenMOSS-Team/OmniVAE">
<img class="hf-icon" src="assets/images/huggingface-logo.svg" alt="" />
<strong>Models</strong>
</a>
</div>
</div>
</section>
<section id="citation" class="section citation-section">
<div class="container">
<div class="section-heading">
<p class="section-label">Citation</p>
<h2>BibTeX</h2>
</div>
<pre><code>@article{zhan2026omnivae,
title = {OmniVAE: An Audio-Video VAE with Cross-Modal Alignment for Joint Generation},
author = {Zhan, Jun and Yang, Chen and Gong, Yitian and Yu, Donghua and Chen, Kuangwei and Zhang, Wenbo and Huang, Kexin and Luo, Qi and Xu, Zhe and Zhu, Ying and Wang, Jin and Zhang, Tengyue and Chen, Qi and Chang, Cheng and Wang, Songlin and Dai, Junqi and Ye, Jiasheng and Yang, Xiaogui and Liang, Tianyi and Peng, Xiangyu and Fei, Zhaoye and Li, Shimin and Cheng, Qinyuan and Chen, Xie and Chen, Xinchi and Qiu, Xipeng},
year = {2026},
url = {https://github.com/OpenMOSS/OmniVAE}
}</code></pre>
</div>
</section>
</main>
<script src="case-data.js?v=9"></script>
</body>
</html>