Skip to content

Commit 2e6f8f2

Browse files
committed
Update LucidNFT project page layout
1 parent 3589791 commit 2e6f8f2

2 files changed

Lines changed: 429 additions & 232 deletions

File tree

LucidNFT/index.html

Lines changed: 125 additions & 85 deletions
Original file line numberDiff line numberDiff line change
@@ -5,62 +5,64 @@
55
<meta charset="UTF-8">
66
<meta name="viewport" content="width=device-width, initial-scale=1.0">
77
<title>LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution</title>
8+
<link rel="stylesheet" href="https://fonts.googleapis.com/css2?family=Roboto:wght@400;500;700&family=Montserrat:wght@400;500;600;700&display=swap">
9+
<link rel="stylesheet" href="https://cdnjs.cloudflare.com/ajax/libs/font-awesome/5.15.3/css/all.min.css">
810
<link rel="stylesheet" href="styles.css">
911
</head>
1012

1113
<body>
12-
<header class="hero" id="top">
13-
<div class="hero-inner">
14+
<div class="sidebar">
15+
<h2 class="sidebar-title">LucidNFT</h2>
16+
<ul>
17+
<li><a href="#paper-info" class="nav-link"><i class="fas fa-file-alt"></i> Paper Information</a></li>
18+
<li><a href="#abstract" class="nav-link"><i class="fas fa-align-left"></i> Abstract</a></li>
19+
<li><a href="#method" class="nav-link"><i class="fas fa-sitemap"></i> Method</a></li>
20+
<li><a href="#lucidlr" class="nav-link"><i class="fas fa-database"></i> LucidLR</a></li>
21+
<li><a href="#results" class="nav-link"><i class="fas fa-chart-line"></i> Results</a></li>
22+
<li><a href="#visual-results" class="nav-link"><i class="fas fa-images"></i> Visual Results</a></li>
23+
<li><a href="#citation" class="nav-link"><i class="fas fa-quote-right"></i> Citation</a></li>
24+
<li><a href="#contact-us" class="nav-link"><i class="fas fa-envelope"></i> Contact Us</a></li>
25+
</ul>
26+
</div>
27+
28+
<div class="content">
29+
<div id="paper-info" class="paper-info">
1430
<h1>LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution</h1>
1531
<p class="tagline">Rollout-group multi-reward fine-tuning for flow-based Real-ISR with LR-referenced faithfulness.</p>
1632

1733
<div class="authors">
18-
<a href="https://feisong123.github.io" target="_blank">Song Fei<sup>1, †</sup></a>
19-
<a href="https://owen718.github.io" target="_blank">Tian Ye<sup>1, †</sup></a>
20-
<a href="https://ephemeral182.github.io" target="_blank">Sixiang Chen<sup>1</sup></a>
21-
<a href="https://ge-xing.github.io" target="_blank">Zhaohu Xing<sup>1</sup></a>
22-
<a href="https://alexlai2860.github.io/" target="_blank">Jianyu Lai<sup>1</sup></a>
23-
<a href="https://sites.google.com/site/indexlzhu/home" target="_blank">Lei Zhu<sup>1, 2, *</sup></a>
34+
<p>
35+
<span class="author"><a href="https://feisong123.github.io" target="_blank">Song Fei<sup>1, †</sup></a></span>,
36+
<span class="author"><a href="https://owen718.github.io" target="_blank">Tian Ye<sup>1, †</sup></a></span>,
37+
<span class="author"><a href="https://ephemeral182.github.io" target="_blank">Sixiang Chen<sup>1</sup></a></span>,
38+
<span class="author"><a href="https://ge-xing.github.io" target="_blank">Zhaohu Xing<sup>1</sup></a></span>,
39+
<span class="author"><a href="https://alexlai2860.github.io/" target="_blank">Jianyu Lai<sup>1</sup></a></span>,
40+
<span class="author"><a href="https://sites.google.com/site/indexlzhu/home" target="_blank">Lei Zhu<sup>1, 2, *</sup></a></span>
41+
</p>
42+
<p>
43+
<span class="affiliation"><sup>1</sup> The Hong Kong University of Science and Technology (Guangzhou)</span>,
44+
<span class="affiliation"><sup>2</sup> The Hong Kong University of Science and Technology</span>
45+
</p>
2446
</div>
2547

26-
<p class="affiliations">
27-
<sup>1</sup>The Hong Kong University of Science and Technology (Guangzhou) ·
28-
<sup>2</sup>The Hong Kong University of Science and Technology
29-
</p>
30-
<p class="notes"><sup></sup>Equal contribution · <sup>*</sup>Corresponding author</p>
31-
32-
<nav class="actions" aria-label="Project links">
33-
<a href="https://arxiv.org/pdf/2603.05947" target="_blank">Paper</a>
34-
<a href="https://github.com/W2GenAI-Lab/LucidNFT" target="_blank">Code</a>
35-
<a href="https://w2genai-lab.github.io/LucidNFT">Project Page</a>
36-
</nav>
48+
<div class="links">
49+
<p><a href="https://arxiv.org/pdf/2603.05947" target="_blank">Paper</a></p>
50+
<p><a href="https://github.com/W2GenAI-Lab/LucidNFT" target="_blank">GitHub</a></p>
51+
<p><a href="https://w2genai-lab.github.io/" target="_blank">Homepage</a></p>
52+
</div>
53+
<p class="notes">† Equal contribution&nbsp;&nbsp;&nbsp;* Corresponding author</p>
3754
</div>
38-
</header>
39-
40-
<nav class="page-nav" aria-label="Page sections">
41-
<a href="#abstract">Abstract</a>
42-
<a href="#motivation">Motivation</a>
43-
<a href="#method">Method</a>
44-
<a href="#lucidlr">LucidLR</a>
45-
<a href="#experiments">Experiments</a>
46-
<a href="#supplement">Supplement</a>
47-
<a href="#citation">Citation</a>
48-
</nav>
4955

50-
<main>
51-
<section id="abstract" class="section">
56+
<section id="abstract" class="info-section">
5257
<h2>Abstract</h2>
5358
<p>
5459
Generative real-world image super-resolution can synthesize visually convincing details from severely degraded low-resolution inputs, yet stochastic sampling makes a critical failure mode hard to avoid: outputs may look sharp but be unfaithful to the LR evidence, exhibiting semantic or structural hallucinations.
5560
</p>
5661
<p>
5762
LucidNFT is a multi-reward RL framework for flow-matching Real-ISR. It introduces LucidConsistency, a degradation-invariant and hallucination-sensitive LR-referenced evaluator trained with content-consistent degradation pools and original-inpainted hard negatives; a decoupled reward normalization strategy that preserves objective-wise contrasts within each LR-conditioned rollout group before fusion; and LucidLR, a large-scale collection of real-world degraded images for robust RL fine-tuning.
5863
</p>
59-
</section>
6064

61-
<section id="motivation" class="section">
62-
<h2>Motivation</h2>
63-
<div class="cards three">
65+
<div class="feature-grid">
6466
<article>
6567
<h3>LR faithfulness is missing</h3>
6668
<p>Without HR references, no-reference perceptual metrics can reward sharp but unsupported details. Real-ISR needs an LR-referenced signal that is robust to degradations and sensitive to hallucination.</p>
@@ -76,7 +78,7 @@ <h3>Real degradations need scale</h3>
7678
</div>
7779
</section>
7880

79-
<section id="method" class="section">
81+
<section id="method" class="info-section">
8082
<h2>Method</h2>
8183
<figure>
8284
<img src="fig/lucidconsistency.png" alt="Overview of LucidConsistency">
@@ -85,7 +87,7 @@ <h2>Method</h2>
8587
</figcaption>
8688
</figure>
8789

88-
<div class="cards three">
90+
<div class="feature-grid">
8991
<article>
9092
<h3>LucidConsistency</h3>
9193
<p>A Qwen3-VL-Embedding-8B backbone with trainable LoRA adapters learns global and native-token representations through pool-based contrastive losses. At inference, it combines global and local LR-SR consistency into an LR-referenced score.</p>
@@ -108,7 +110,7 @@ <h3>LucidNFT fine-tuning</h3>
108110
</figure>
109111
</section>
110112

111-
<section id="lucidlr" class="section">
113+
<section id="lucidlr" class="info-section">
112114
<h2>LucidLR Dataset</h2>
113115
<p>
114116
LucidLR is a 20K-image real-world low-quality dataset collected from Wikimedia Commons through its official API. Images are gathered from public low-quality and blurred-image categories, filtered from an approximately 22K-image raw pool with NSFW classification, corrupted-file removal, and manual review.
@@ -118,7 +120,7 @@ <h2>LucidLR Dataset</h2>
118120
<figcaption>Representative LucidLR samples with diverse real-world degradations, used as LR inputs for RL fine-tuning.</figcaption>
119121
</figure>
120122

121-
<div class="table-wrap">
123+
<div class="table-wrap compact-table">
122124
<table>
123125
<caption>Comparison of representative real-world datasets used in Real-ISR.</caption>
124126
<thead>
@@ -140,8 +142,8 @@ <h2>LucidLR Dataset</h2>
140142
</div>
141143
</section>
142144

143-
<section id="experiments" class="section">
144-
<h2>Experiments</h2>
145+
<section id="results" class="info-section">
146+
<h2>Results</h2>
145147
<p>
146148
Experiments evaluate LucidNFT on two flow-based Real-ISR models, LucidFlux and DiT4SR. All methods are evaluated at 1024 x 1024 output resolution with 4x upscaling. The paper reports eight no-reference quality metrics and LucidConsistency as an LR-referenced consistency score without HR ground truth.
147149
</p>
@@ -153,7 +155,7 @@ <h2>Experiments</h2>
153155
</figcaption>
154156
</figure>
155157

156-
<div class="table-wrap wide">
158+
<div class="table-wrap wide-table">
157159
<table>
158160
<caption>Quantitative comparison on RealLQ250, DRealSR, and RealSR. Higher is better except NIQE.</caption>
159161
<thead>
@@ -205,19 +207,8 @@ <h2>Experiments</h2>
205207
</table>
206208
</div>
207209

208-
<div class="figure-grid">
209-
<figure>
210-
<img src="visualization/visualization_1.png" alt="Visual comparison on RealLQ250">
211-
<figcaption>Main qualitative comparison on RealLQ250.</figcaption>
212-
</figure>
213-
<figure>
214-
<img src="visualization/visualization_2.png" alt="Additional qualitative comparison on RealLQ250">
215-
<figcaption>Additional qualitative comparisons on RealLQ250.</figcaption>
216-
</figure>
217-
</div>
218-
219210
<div class="table-grid">
220-
<div class="table-wrap">
211+
<div class="table-wrap compact-table">
221212
<table>
222213
<caption>Human-aligned LR-faithfulness evaluation on RealSR.</caption>
223214
<thead>
@@ -234,7 +225,7 @@ <h2>Experiments</h2>
234225
</table>
235226
</div>
236227

237-
<div class="table-wrap">
228+
<div class="table-wrap compact-table">
238229
<table>
239230
<caption>Ablation study on RealLQ250 using LucidFlux.</caption>
240231
<thead>
@@ -253,46 +244,95 @@ <h2>Experiments</h2>
253244
</div>
254245
</section>
255246

256-
<section id="supplement" class="section">
257-
<h2>Supplemental Figures</h2>
258-
<div class="figure-grid">
259-
<figure>
260-
<img src="fig/human_interface.png" alt="Human preference annotation interface">
261-
<figcaption>Human preference annotation interface for LR-faithfulness evaluation.</figcaption>
262-
</figure>
263-
<figure>
264-
<img src="fig/degradation_primary.png" alt="Primary degradation distribution">
265-
<figcaption>Primary degradation distribution across LucidLR, RealLQ250, RealSR, and DRealSR.</figcaption>
266-
</figure>
267-
<figure>
268-
<img src="fig/degradation_occurrence.png" alt="Occurrence frequency of degradation categories">
269-
<figcaption>Occurrence frequency of degradation categories. LucidLR provides broader degradation coverage and a richer long-tail distribution.</figcaption>
270-
</figure>
271-
<figure>
272-
<img src="fig/training_curves_lucidflux_ablation.png" alt="Optimization curves under different reward formulations">
273-
<figcaption>Optimization curves under different reward formulations on LucidFlux.</figcaption>
274-
</figure>
247+
<section id="visual-results" class="info-section">
248+
<h2>Visual Results</h2>
249+
<div class="sample-gallery">
250+
<article>
251+
<h3>Sample 016</h3>
252+
<div class="sample-columns">
253+
<figure>
254+
<img src="gallery/016_lq.jpg" alt="Sample 016 low-quality input">
255+
<figcaption>LR input</figcaption>
256+
</figure>
257+
<figure>
258+
<img src="gallery/016_lucidflux.jpg" alt="Sample 016 LucidFlux result">
259+
<figcaption>LucidFlux</figcaption>
260+
</figure>
261+
<figure>
262+
<img src="gallery/016_lucidnft.jpg" alt="Sample 016 LucidNFT result">
263+
<figcaption>LucidNFT</figcaption>
264+
</figure>
265+
</div>
266+
</article>
267+
<article>
268+
<h3>Sample 137</h3>
269+
<div class="sample-columns">
270+
<figure>
271+
<img src="gallery/137_lq.jpg" alt="Sample 137 low-quality input">
272+
<figcaption>LR input</figcaption>
273+
</figure>
274+
<figure>
275+
<img src="gallery/137_lucidflux.jpg" alt="Sample 137 LucidFlux result">
276+
<figcaption>LucidFlux</figcaption>
277+
</figure>
278+
<figure>
279+
<img src="gallery/137_lucidnft.jpg" alt="Sample 137 LucidNFT result">
280+
<figcaption>LucidNFT</figcaption>
281+
</figure>
282+
</div>
283+
</article>
284+
<article>
285+
<h3>Sample 182</h3>
286+
<div class="sample-columns">
287+
<figure>
288+
<img src="gallery/182_lq.png" alt="Sample 182 low-quality input">
289+
<figcaption>LR input</figcaption>
290+
</figure>
291+
<figure>
292+
<img src="gallery/182_lucidflux.jpg" alt="Sample 182 LucidFlux result">
293+
<figcaption>LucidFlux</figcaption>
294+
</figure>
295+
<figure>
296+
<img src="gallery/182_lucidnft.jpg" alt="Sample 182 LucidNFT result">
297+
<figcaption>LucidNFT</figcaption>
298+
</figure>
299+
</div>
300+
</article>
301+
<article>
302+
<h3>Sample 223</h3>
303+
<div class="sample-columns">
304+
<figure>
305+
<img src="gallery/223_lq.png" alt="Sample 223 low-quality input">
306+
<figcaption>LR input</figcaption>
307+
</figure>
308+
<figure>
309+
<img src="gallery/223_lucidflux.jpg" alt="Sample 223 LucidFlux result">
310+
<figcaption>LucidFlux</figcaption>
311+
</figure>
312+
<figure>
313+
<img src="gallery/223_lucidnft.jpg" alt="Sample 223 LucidNFT result">
314+
<figcaption>LucidNFT</figcaption>
315+
</figure>
316+
</div>
317+
</article>
275318
</div>
276319
</section>
277320

278-
<section id="citation" class="section">
279-
<h2>Conclusion</h2>
280-
<p>
281-
LucidNFT aligns stochastic restorations not only toward perceptual realism, but also toward LR-conditioned faithfulness. Experiments across two flow-based Real-ISR backbones show improved perceptual quality while generally maintaining LR-referenced consistency, suggesting a practical route toward more reliable generative restoration.
282-
</p>
321+
<section id="citation" class="paper-citation">
283322
<h2>Citation</h2>
284323
<pre><code>@article{fei2026lucidnft,
285324
title={LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution},
286325
author={Fei, Song and Ye, Tian and Chen, Sixiang and Xing, Zhaohu and Lai, Jianyu and Zhu, Lei},
287-
journal={arXiv preprint},
326+
journal={arXiv},
288327
year={2026}
289328
}</code></pre>
290329
</section>
291-
</main>
292330

293-
<footer>
294-
<p>For questions or collaboration, contact sfei285@connect.hkust-gz.edu.cn, tye610@connect.hkust-gz.edu.cn, or leizhu@hkust-gz.edu.cn.</p>
295-
</footer>
331+
<section id="contact-us" class="contact-us">
332+
<h2>Contact Us</h2>
333+
<p>For questions or collaboration, contact sfei285@connect.hkust-gz.edu.cn, tye610@connect.hkust-gz.edu.cn, or leizhu@hkust-gz.edu.cn.</p>
334+
</section>
335+
</div>
296336
</body>
297337

298338
</html>

0 commit comments

Comments
 (0)