-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
505 lines (463 loc) · 22.5 KB
/
Copy pathindex.html
File metadata and controls
505 lines (463 loc) · 22.5 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
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<!-- ========= Primary Meta Tags ========= -->
<meta name="title" content="Mirai: Autoregressive Visual Generation Needs Foresight">
<meta name="description" content="Mirai is a foresight alignment framework for autoregressive image generation that injects future signals into training, improving global coherence and convergence.">
<meta name="keywords" content="Mirai, autoregressive image generation, visual autoregressive models, foresight alignment, LlamaGen, ImageNet, generative modeling, computer vision, deep learning">
<meta name="author" content="Yonghao Yu, Lang Huang, Zerun Wang, Runyi Li, Toshihiko Yamasaki">
<meta name="robots" content="index, follow">
<meta name="language" content="English">
<!-- ========= Open Graph / Facebook ========= -->
<meta property="og:type" content="article">
<meta property="og:site_name" content="Mirai Project Page">
<meta property="og:title" content="Mirai: Autoregressive Visual Generation Needs Foresight">
<meta property="og:description" content="Mirai is a foresight alignment framework for autoregressive image generation that injects future signals into training, improving global coherence and convergence.">
<meta property="og:url" content="https://y0urOy.github.io/Mirai/">
<meta property="og:image" content="https://y0urOy.github.io/Mirai/static/images/social_preview.png">
<meta property="og:image:width" content="1200">
<meta property="og:image:height" content="630">
<meta property="og:image:alt" content="Mirai: Autoregressive Visual Generation Needs Foresight - teaser">
<meta property="article:published_time" content="2025-12-01T00:00:00.000Z">
<meta property="article:author" content="Yonghao Yu">
<meta property="article:section" content="Research">
<meta property="article:tag" content="autoregressive image generation">
<meta property="article:tag" content="foresight alignment">
<!-- ========= Twitter（暂时不用）========= -->
<!--
<meta name="twitter:card" content="summary_large_image">
<meta name="twitter:site" content="@YOUR_TWITTER_HANDLE">
<meta name="twitter:creator" content="@AUTHOR_TWITTER_HANDLE">
<meta name="twitter:title" content="Mirai: Autoregressive Visual Generation Needs Foresight">
<meta name="twitter:description" content="Mirai is a foresight alignment framework for autoregressive image generation that injects future signals into training, improving global coherence and convergence.">
<meta name="twitter:image" content="https://y0urOy.github.io/Mirai/static/images/social_preview.png">
<meta name="twitter:image:alt" content="Mirai: Autoregressive Visual Generation Needs Foresight - teaser">
-->
<!-- ========= Academic / Research Specific ========= -->
<meta name="citation_title" content="Mirai: Autoregressive Visual Generation Needs Foresight">
<meta name="citation_author" content="Yu, Yonghao">
<meta name="citation_author" content="Huang, Lang">
<meta name="citation_author" content="Wang, Zerun">
<meta name="citation_author" content="Li, Runyi">
<meta name="citation_author" content="Yamasaki, Toshihiko">
<!-- <meta name="citation_publication_date" content="2026">
<meta name="citation_conference_title" content="CVPR 2026 (under review)"> -->
<!-- 如果暂时没有公开 PDF，这里保留占位也没问题 -->
<meta name="citation_pdf_url" content="https://y0urOy.github.io/Mirai/static/pdfs/mirai_cvpr2026.pdf">
<!-- ========= Additional SEO ========= -->
<meta name="theme-color" content="#2563eb">
<meta name="msapplication-TileColor" content="#2563eb">
<meta name="apple-mobile-web-app-capable" content="yes">
<meta name="apple-mobile-web-app-status-bar-style" content="default">
<!-- ========= Preconnect ========= -->
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link rel="preconnect" href="https://ajax.googleapis.com">
<link rel="preconnect" href="https://documentcloud.adobe.com">
<link rel="preconnect" href="https://cdn.jsdelivr.net">
<!-- 页面标题 -->
<title>Mirai: Autoregressive Visual Generation Needs Foresight</title>
<!-- Favicon -->
<link rel="icon" type="image/x-icon" href="static/images/favicon.ico">
<link rel="apple-touch-icon" href="static/images/favicon.ico">
<!-- Critical CSS -->
<link rel="stylesheet" href="static/css/bulma.min.css">
<link rel="stylesheet" href="static/css/index.css">
<!-- Non-critical CSS -->
<link rel="preload" href="static/css/bulma-carousel.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="static/css/bulma-slider.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="static/css/fontawesome.all.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<link rel="preload" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css" as="style" onload="this.onload=null;this.rel='stylesheet'">
<noscript>
<link rel="stylesheet" href="static/css/bulma-carousel.min.css">
<link rel="stylesheet" href="static/css/bulma-slider.min.css">
<link rel="stylesheet" href="static/css/fontawesome.all.min.css">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/gh/jpswalsh/academicons@1/css/academicons.min.css">
</noscript>
<!-- Fonts -->
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700;800&display=swap" rel="stylesheet">
<!-- JS -->
<script defer src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script defer src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script defer src="static/js/bulma-carousel.min.js"></script>
<script defer src="static/js/bulma-slider.min.js"></script>
<script defer src="static/js/index.js"></script>
<!-- ========= Structured Data for Academic Papers ========= -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "ScholarlyArticle",
"headline": "Mirai: Autoregressive Visual Generation Needs Foresight",
"description": "Mirai is a foresight alignment framework for autoregressive image generation that injects future signals into training, improving global coherence and convergence.",
"author": [
{
"@type": "Person",
"name": "Yonghao Yu",
"affiliation": {
"@type": "Organization",
"name": "The University of Tokyo"
}
},
{
"@type": "Person",
"name": "Lang Huang",
"affiliation": {
"@type": "Organization",
"name": "National Institute of Informatics"
}
},
{
"@type": "Person",
"name": "Zerun Wang",
"affiliation": {
"@type": "Organization",
"name": "The University of Tokyo"
}
},
{
"@type": "Person",
"name": "Runyi Li",
"affiliation": {
"@type": "Organization",
"name": "Peking University"
}
},
{
"@type": "Person",
"name": "Toshihiko Yamasaki",
"affiliation": {
"@type": "Organization",
"name": "The University of Tokyo"
}
}
],
"datePublished": "2025-12-01",
"publisher": {
"@type": "Organization",
"name": ""
},
"url": "https://y0urOy.github.io/Mirai/",
"image": "https://y0urOy.github.io/Mirai/static/images/social_preview.png",
"keywords": [
"autoregressive image generation",
"visual autoregressive models",
"foresight alignment",
"LlamaGen",
"ImageNet",
"deep generative models"
],
"abstract": "Mirai is a foresight alignment framework for autoregressive visual generators. By injecting training signals from future tokens and aligning them on the 2D image grid, Mirai significantly improves global coherence and accelerates convergence over standard next-token training.",
"isAccessibleForFree": true,
"license": "https://creativecommons.org/licenses/by/4.0/",
"mainEntity": {
"@type": "WebPage",
"@id": "https://y0urOy.github.io/Mirai/"
}
}
</script>
<!-- ========= Organization Structured Data ========= -->
<script type="application/ld+json">
{
"@context": "https://schema.org",
"@type": "Organization",
"name": "Mirai Project",
"url": "https://y0urOy.github.io/Mirai/",
"logo": "https://y0urOy.github.io/Mirai/static/images/favicon.ico",
"sameAs": [
"https://github.com/y0urOy"
]
}
</script>
</head>
<body>
<!-- Scroll to Top Button -->
<button class="scroll-to-top" onclick="scrollToTop()" title="Scroll to top" aria-label="Scroll to top">
<i class="fas fa-chevron-up"></i>
</button>
<!-- More Works 下拉：暂时不用 -->
<!--
<div class="more-works-container">
...
</div>
-->
<main id="main-content">
<!-- ========== 顶部标题 + 作者 + 按钮区域 ========== -->
<section class="hero">
<div class="hero-body" style="padding-top: 3rem; padding-bottom: 1.0rem;">
<div class="container is-max-desktop">
<div class="columns is-centered">
<div class="column has-text-centered">
<h1 class="title is-1 publication-title">
Mirai: Autoregressive Visual Generation Needs Foresight
</h1>
<!-- 作者 + 机构编号 + * 对应作者 -->
<div class="is-size-5 publication-authors">
<span class="author-block">Yonghao Yu<sup>1</sup></span>,
<span class="author-block">Lang Huang<sup>2*</sup></span>,
<span class="author-block">Zerun Wang<sup>1</sup></span>,
<span class="author-block">Runyi Li<sup>3</sup></span>,
<span class="author-block">Toshihiko Yamasaki<sup>1</sup></span>
</div>
<div class="is-size-5 publication-authors" style="margin-top: 0.4rem;">
<span class="author-block">
<sup>1</sup> The University of Tokyo ·
<sup>2</sup> National Institute of Informatics ·
<sup>3</sup> Peking University
</span>
<span class="author-block">
<small><sup>*</sup> Corresponding author</small>
</span>
</div>
<!-- 顶部按钮：Code + arXiv -->
<div class="column has-text-centered" style="margin-top: 0.4rem;">
<div class="publication-links">
<!-- Code -->
<span class="link-block">
<a href="https://github.com/y0urOy/Mirai-Code" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fab fa-github"></i>
</span>
<span>Code</span>
</a>
</span>
<!-- arXiv（有号之后替换链接） -->
<span class="link-block">
<a href="https://arxiv.org/abs/2601.14671" target="_blank"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="ai ai-arxiv"></i>
</span>
<span>arXiv</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- ========== Teaser ========== -->
<section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body has-text-centered" style="padding-top: 0.8rem; padding-bottom: 3rem;">
<img src="static/images/teaser.png"
alt="Mirai teaser figure"
style="max-width: 100%; border-radius: 8px;">
<h2 class="subtitle has-text-centered" style="margin-top: 1rem;">
<strong>Left:</strong> Comparison between LlamaGen-B and Mirai after 300 epochs.
<strong>Right:</strong> Mirai significantly accelerates training convergence on ImageNet.
</h2>
</div>
</div>
</section>
<!-- ========== Abstract ========== -->
<section class="section hero is-light">
<div class="container is-max-desktop">
<div class="columns is-centered has-text-centered">
<div class="column is-four-fifths">
<h2 class="title is-3" style="margin-bottom: 1rem; padding-bottom: 0;">
Abstract
</h2>
<div class="content has-text-justified" style="line-height: 1.5;">
<p style="margin-bottom: 0.5rem;">
Autoregressive (AR) visual generators model images as sequences of discrete tokens and are trained with next-token likelihood.
This strict causality supervision optimizes each step only by its immediate next token, which diminishes global coherence and slows convergence.
</p>
<p style="margin-bottom: 0.5rem;">
We ask whether <strong>foresight</strong>—training signals that originate from later tokens—can help AR visual generation.
Through controlled diagnostics across injection level, spatial layout, and foresight source, we find that aligning foresight to AR models' internal representation on the 2D image grids is crucial for effective causality modeling.
</p>
<p style="margin-bottom: 0.5rem;">
We instantiate this insight with <strong>Mirai</strong> (meaning “future” in Japanese), a general framework that injects future information
into AR training without architecture changes or inference-time overhead. <strong>Mirai-E</strong> uses explicit foresight from multiple future positions of unidirectional representations, whereas <strong>Mirai-I</strong> everages implicit foresight from matched bidirectional representations.
Extensive experiments on ImageNet show that Mirai both accelerates convergence and improves generation quality.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- ========== Method Overview ========== -->
<section class="hero is-small">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered"
style="margin-bottom: 1rem; padding-bottom: 0;">
Method Overview
</h2>
<div class="columns is-centered">
<div class="column is-10">
<img src="static/images/mirai_figures.png"
alt="Mirai framework overview"
style="max-width: 100%; border-radius: 8px;">
<p class="subtitle has-text-centered" style="margin-top: 0.75rem;">
Mirai aligns the AR model's <strong>internal representations</strong> with the foresight from either <strong>bidirectional</strong> or <strong>unidirectional</strong> foresight encoder in a <strong>2D grid</strong>.
</p>
<p class="subtitle" style="margin-top: 0.25rem; margin-bottom: 0.25px; text-align: center; word-break: break-word; margin-left: auto; margin-right: auto; line-height: 1.4;">
Depending on the source of the foresight, Mirai admits two instantiations:
</p>
<p class="subtitle" style="margin-top: 0.25px; margin-bottom: 0px; font-size: 1.1rem; text-align: justify; line-height: 1.4; word-break: break-word; margin-left: auto; margin-right: auto;">
<strong>Mirai-E</strong> provides explicit, position–indexed foresight from the unidirectional AR model’s own Exponential Moving Average (EMA), aligning internal state to the foresights at a small set of nearby future locations.
</p>
<div class="columns is-centered" style="margin: 0;">
<div class="column is-8 has-text-centered" style="padding: 0;">
<img src="static/images/parrot_resize.gif"
alt="Mirai-E alignment"
style="width: 65%; border-radius: 8px; display: block; margin: 0 auto;">
<p style="
margin: 0.1rem 0 1.2rem 0;
font-size: 1.15rem;
font-weight: 500;
">
Left: Mirai-E alignment visualization. Right: Strict casual.
</p>
</div>
</div>
<p class="subtitle" style="margin-top: 0px; font-size: 1.1rem; text-align: justify; line-height: 1.4; word-break: break-word; margin-left: auto; margin-right: auto;">
<strong>Mirai-I</strong> supplies implicit, context–aggregating foresight by aligning internal states to features from a frozen bidirectional encoder at matched spatial locations.
</p>
<!-- <p class="subtitle" style="margin-top: 0rem; font-size: 1.1rem; text-align: justify; word-break: break-word; margin-left: auto; margin-right: auto; line-height: 1.4;">
At test time, the additional alignment components are removed; decoding remains token-by-token, strictly causal, and identical in computational cost to the standard AR model. -->
</h2>
</div>
</div>
</div>
</div>
</section>
<!-- ========== Results / Samples ========== -->
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered"
style="margin-bottom: 1rem; padding-bottom: 0;">
System-Level Comparison
</h2>
<div class="columns is-centered">
<div class="column is-10 has-text-centered">
<img src="static/images/system_results.png"
alt="Generated samples from Mirai"
style="max-width: 100%; border-radius: 8px;">
<h2 class="subtitle has-text-centered" style="margin-top: 1rem;">
System-Level Comparison on ImageNet 256×256. ↓ and ↑ indicate whether lower or higher values are better, respectively.
</h2>
</div>
</div>
</div>
</div>
</section>
<!-- ========== Results / Samples ========== -->
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered"
style="margin-bottom: 1rem; padding-bottom: 0;">
FID Comparisons
</h2>
<div class="columns is-centered">
<div class="column is-10 has-text-centered">
<img src="static/images/curve.png"
alt="Generated samples from Mirai"
style="max-width: 100%; border-radius: 8px;">
<h2 class="subtitle has-text-centered" style="margin-top: 1rem;">
FID comparisons between Mirai with vanilla LlamaGen across different model sizes and epochs on ImageNet 256×256.
</h2>
</div>
</div>
</div>
</div>
</section>
<!-- ========== Results / Samples ========== -->
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered"
style="margin-bottom: 1rem; padding-bottom: 0;">
Internal Representation Visualization
</h2>
<div class="columns is-centered">
<div class="column is-10 has-text-centered">
<img src="static/images/tsne.png"
alt="Generated samples from Mirai"
style="max-width: 100%; border-radius: 8px;">
<h2 class="subtitle has-text-centered" style="margin-top: 1rem;">
Visualization of layer-8 internal representations on the 2D token grid. Each token’s 2D t-SNE embedding is mapped to a
color (with the Color Map at bottom left) and plotted at its original grid location. Smooth color fields indicate 2D-structured representations;
the red rectangle in LlamaGen-B highlights abrupt color changes where spatial structure breaks down.
</h2>
</div>
</div>
</div>
</div>
</section>
<!-- ========== Results / Samples ========== -->
<section class="hero is-small is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<h2 class="title is-3 has-text-centered"
style="margin-bottom: 1rem; padding-bottom: 0;">
ImageNet Results
</h2>
<div class="columns is-centered">
<div class="column is-10 has-text-centered">
<img src="static/images/sample.png"
alt="Generated samples from Mirai"
style="max-width: 100%; border-radius: 8px;">
<h2 class="subtitle has-text-centered" style="margin-top: 1rem;">
Generated samples on ImageNet 256×256 from LlamaGen-XL with Mirai-I. Mirai improves global
structure, object integrity, and perceptual quality compared to the AR baseline.
</h2>
</div>
</div>
</div>
</div>
</section>
<!-- ========== BibTeX ========== -->
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<div class="bibtex-header">
<h2 class="title">BibTeX</h2>
<button class="copy-bibtex-btn" onclick="copyBibTeX()" title="Copy BibTeX to clipboard">
<i class="fas fa-copy"></i>
<span class="copy-text">Copy</span>
</button>
</div>
<pre id="bibtex-code"><code>@article{yu2026mirai,
title={Mirai: Autoregressive Visual Generation Needs Foresight},
author={Yu, Yonghao and Huang, Lang and Wang, Zerun and Li, Runyi and Yamasaki, Toshihiko},
journal={arXiv preprint arXiv:2601.14671},
year={2026}
}
</code></pre>
</div>
</section>
</main>
<!-- ========= Footer ========= -->
<footer class="footer">
<div class="container">
<div class="columns is-centered">
<div class="column is-8">
<div class="content">
<p>
This page was built using the
<a href="https://github.com/eliahuhorwitz/Academic-project-page-template" target="_blank">Academic Project Page Template</a>,
which was adapted from the <a href="https://nerfies.github.io" target="_blank">Nerfies</a> project page.
You are free to borrow the source code of this website; we just ask that you link back to this page in the footer.
<br>
This website is licensed under a
<a rel="license" href="http://creativecommons.org/licenses/by-sa/4.0/" target="_blank">
Creative Commons Attribution-ShareAlike 4.0 International License
</a>.
</p>
</div>
</div>
</div>
</div>
</footer>
<!-- Statcounter tracking code（不用就保持空） -->
</body>
</html>