Skip to content

Commit 4912066

Browse files
hao-pthao-pt
committed
update mobile rendering
1 parent 5cf1442 commit 4912066

File tree

3 files changed

+46
-72
lines changed

3 files changed

+46
-72
lines changed

‎index.html

Lines changed: 41 additions & 72 deletions
Original file line numberDiff line numberDiff line change
@@ -108,7 +108,7 @@
108108
<div class="container is-max-desktop">
109109
<div class="columns is-centered">
110110
<div class="column has-text-centered">
111-
<h1 class="title is-1 publication-title">DiMSUM <img src="./static/images/dimsum_icon.png" class="logo" width=5.5% />: <span style="color:red;">Di</span>ffusion <span style="color:red;">M</span>amba - A <span style="color:red;">S</span>calable and <span style="color:red;">U</span>nified
111+
<h1 class="title is-1 publication-title">DiMSUM <img src="./static/images/dimsum_icon.png" class="logo" width="50px" />: <span style="color:red;">Di</span>ffusion <span style="color:red;">M</span>amba - A <span style="color:red;">S</span>calable and <span style="color:red;">U</span>nified
112112
Spatial-Frequency <span style="color:red;">M</span>ethod for Image Generation</h1>
113113
<div class="is-size-5 publication-authors">
114114
<span class="author-block">
@@ -124,10 +124,10 @@ <h1 class="title is-1 publication-title">DiMSUM <img src="./static/images/dimsum
124124
<a href="https://viethoang1512.github.io/">Hoang Phan</a><sup>4</sup>&emsp;
125125
</span>
126126
<span class="author-block">
127-
<a href="https://people.cs.rutgers.edu/~dnm/">Dimitris N. Metaxas</a><sup>3</sup>&emsp;
127+
<a href="https://people.cs.rutgers.edu/~dnm/">Dimitris N. Metaxas</a><sup>2</sup>&emsp;
128128
</span>
129129
<span class="author-block">
130-
<a href="https://sites.google.com/site/anhttranusc/">Anh Tran</a><sup>1</sup>
130+
<a href="https://scholar.google.com/citations?user=FYZ5ODQAAAAJ">Anh Tran</a><sup>1</sup>
131131
</span>
132132
</div>
133133

@@ -524,11 +524,10 @@ <h2 class="title is-3">Unconditional Generation</h2>
524524

525525
<!-- Motivation -->
526526
<h2 class="title is-3">Why is scanning in frequency space helpful?</h2>
527-
<div class="columns is-centered">
527+
<div class="columns is-centered" style="text-align: center;">
528528
<img src="./static/images/wavelet_vs_spatial_window.png" width="60%" class="scanning" />
529529
</div>
530530
<div class="content has-text-justified">
531-
532531
<p>
533532
Previous state-space models, particularly in processing visual data, failed to effectively address the design choice of scanning order due to their exclusive reliance on spatial processing, neglecting crucial long-range relations in the frequency spectrum.
534533
We propose a novel approach that integrates frequency scanning with the conventional spatial scanning mechanism.
@@ -595,88 +594,58 @@ <h3 class="title is-4">Globally-shared Transformer Block</h3>
595594
<div class="column is-full-width">
596595
<h2 class="title is-3">Results</h2>
597596
</div>
598-
<div class="columns is-centered">
599-
<div class="column">
600-
<div class="content">
601-
<figure>
602-
<img src="./static/images/celeb256.jpg" class="interpolation-image" width=45% style="margin-right: 60px;" />
603-
<img src="./static/images/celeb512.jpg" class="interpolation-image" width=45% />
604-
<figcaption>Figure 1. Unconditional generation on CelebA HQ</figcaption>
605-
</figure>
606-
607-
</div>
608-
</div>
609-
</div>
610-
<br>
597+
<div class="columns is-centered" style="text-align: center;">
598+
<table>
599+
<tr>
600+
<td><img src="./static/images/celeb256.jpg" width="90%" /> </td>
601+
<td> <img src="./static/images/celeb512.jpg" width="90%" /> </td>
602+
</tr>
603+
<caption style="text-align: center; color: black">Figure 1. Unconditional generation on CelebA HQ 256 & 512</caption>
604+
</table>
611605
</div>
612-
<div class="columns is-centered">
613-
<div class="column" style="margin-left: 13%;">
614-
<div class="content">
615-
<figure>
616-
<img src="./static/images/church256.jpg" class="interpolation-image" width=70% />
617-
<figcaption>Figure 2. Unconditional generation on LSUN Church</figcaption>
618-
</figure>
619-
</div>
620-
<!-- <div class="content">
621-
<figure>
622-
<img src="./static/images/imnet.jpg" class="interpolation-image" width=35% />
623-
<figcaption>Figure 3. Class-conditional generation on ImageNet1k 256</figcaption>
624-
</figure>
625-
</div> -->
606+
<div class="columns is-centered" style="text-align: center;">
607+
<div class="column">
608+
<figure >
609+
<img src="./static/images/training_convergence.png" class="interpolation-image" width="70%"/>
610+
<figcaption>Figure 2. Training convergence on CelebA HQ 256.</figcaption>
611+
</figure>
626612
</div>
627613
<div class="column">
628-
<div class="content" style="margin-left: -42%;">
629-
<figure>
630-
<img src="./static/images/imnet.jpg" class="interpolation-image" width=60% />
631-
<figcaption>Figure 3. Class-conditional generation on ImageNet1k 256</figcaption>
632-
</figure>
633-
</div>
614+
<figure>
615+
<img src="./static/images/church256.jpg" class="interpolation-image" width="90%" />
616+
<figcaption>Figure 3. Unconditional generation on LSUN Church</figcaption>
617+
</figure>
634618
</div>
635-
<!-- <div class="column">
636-
<div class="content">
637-
<figure>
638-
<img src="./static/images/training_convergence.png" class="interpolation-image" width=35% />
639-
<figcaption style="margin-left: 400px; margin-right: 400px;">
640-
Figure 4. Training convergence on CelebA HQ 256.
641-
Our method achieves faster training convergence, requiring fewer than half the training epochs compared to other diffusion models, while delivering a more stable training curve.
642-
</figcaption>
643-
</figure>
644-
</div>
645-
</div> -->
619+
</div>
620+
<div class="columns is-centered" style="text-align: center;">
646621
</div>
647622
<div class="columns is-centered">
648623
<div class="column">
649624
<div class="content">
650625
<figure>
651-
<img src="./static/images/training_convergence.png" class="interpolation-image" width=35% />
652-
<figcaption style="margin-left: 400px; margin-right: 400px;">
653-
Figure 4. Training convergence on CelebA HQ 256.
654-
Our method achieves faster training convergence, requiring fewer than half the training epochs compared to other diffusion models, while delivering a more stable training curve.
655-
</figcaption>
626+
<img src="./static/images/imnet.jpg" class="interpolation-image" width=60% />
627+
<figcaption>Figure 4. Class-conditional generation on ImageNet1k 256</figcaption>
656628
</figure>
657629
</div>
658630
</div>
659631
</div>
660632
</section>
661-
662-
<!-- <section class="section" id="bound">
663-
<div class="container is-max-desktop content">
664-
<h2 class="title">Theoretical analysis: Bounding estimation error</h2>
665-
<div class="content has-text-justified">
666-
We have shown that minimizing the FM objective on latent space controls the Wasserstein distance between the
667-
target density \( p_0 \) and the reconstructed density \( \hat{p}_0 \), which coincides with Fréchet inception
668-
distance (FID), a common metric for image generation.
669-
This means that our latent flow matching is guaranteed to control this metric, given reasonable estimation of \(
670-
\hat{v}(\mathbf{z}_t, t) \). Nonetheless, the analysis also suggests that the quality of latent flow matching
671-
depends on the constants that define the expressivity of the decoders and encoders, which has been observed in
672-
prior research on generative modeling in latent space.
673-
674-
<img src="./static/images/bound.jpeg" class="interpolation-image" />
675-
676-
633+
634+
635+
<section class="section" id="speed">
636+
<div class="container is-max-desktop">
637+
<div class="column is-full-width">
638+
<h2 class="title is-3">Speed</h2>
677639
</div>
678-
</div>
679-
</section> -->
640+
<div class="container is-max-desktop content" style="text-align: center;">
641+
<div>
642+
<img src="./static/images/speed.jpg" class="interpolation-image" width="80%"/>
643+
</div>
644+
<div class="content has-text-justified">
645+
The speed gap between our method and DiT widens as the input resolution increases, highlighting the efficiency of our method for high-resolution synthesis.
646+
</div>
647+
</div>
648+
</section>
680649

681650
<!-- <section class="section" id="Related">
682651
<div class="container is-max-desktop content">

‎static/css/index.css

Lines changed: 5 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -251,4 +251,9 @@ background-color: #F2E34C;
251251
--mask-image: repeating-linear-gradient(135deg, #000 0px, #000 1px, transparent 1px, transparent 6px);
252252
-webkit-mask-image: var(--mask-image);
253253
mask-image: var(--mask-image);
254+
}
255+
256+
@media (max-width:600px) {
257+
.about-us-text,
258+
.about-us-images { width:100% }
254259
}

‎static/images/speed.jpg

74.6 KB
Loading

0 commit comments

Comments
 (0)