You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
55
60
</p>
56
61
<p>
57
62
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.
58
63
</p>
59
-
</section>
60
64
61
-
<sectionid="motivation" class="section">
62
-
<h2>Motivation</h2>
63
-
<divclass="cards three">
65
+
<divclass="feature-grid">
64
66
<article>
65
67
<h3>LR faithfulness is missing</h3>
66
68
<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>
76
78
</div>
77
79
</section>
78
80
79
-
<sectionid="method" class="section">
81
+
<sectionid="method" class="info-section">
80
82
<h2>Method</h2>
81
83
<figure>
82
84
<imgsrc="fig/lucidconsistency.png" alt="Overview of LucidConsistency">
@@ -85,7 +87,7 @@ <h2>Method</h2>
85
87
</figcaption>
86
88
</figure>
87
89
88
-
<divclass="cards three">
90
+
<divclass="feature-grid">
89
91
<article>
90
92
<h3>LucidConsistency</h3>
91
93
<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>
108
110
</figure>
109
111
</section>
110
112
111
-
<sectionid="lucidlr" class="section">
113
+
<sectionid="lucidlr" class="info-section">
112
114
<h2>LucidLR Dataset</h2>
113
115
<p>
114
116
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>
118
120
<figcaption>Representative LucidLR samples with diverse real-world degradations, used as LR inputs for RL fine-tuning.</figcaption>
119
121
</figure>
120
122
121
-
<divclass="table-wrap">
123
+
<divclass="table-wrap compact-table">
122
124
<table>
123
125
<caption>Comparison of representative real-world datasets used in Real-ISR.</caption>
124
126
<thead>
@@ -140,8 +142,8 @@ <h2>LucidLR Dataset</h2>
140
142
</div>
141
143
</section>
142
144
143
-
<sectionid="experiments" class="section">
144
-
<h2>Experiments</h2>
145
+
<sectionid="results" class="info-section">
146
+
<h2>Results</h2>
145
147
<p>
146
148
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.
147
149
</p>
@@ -153,7 +155,7 @@ <h2>Experiments</h2>
153
155
</figcaption>
154
156
</figure>
155
157
156
-
<divclass="table-wrap wide">
158
+
<divclass="table-wrap wide-table">
157
159
<table>
158
160
<caption>Quantitative comparison on RealLQ250, DRealSR, and RealSR. Higher is better except NIQE.</caption>
159
161
<thead>
@@ -205,19 +207,8 @@ <h2>Experiments</h2>
205
207
</table>
206
208
</div>
207
209
208
-
<divclass="figure-grid">
209
-
<figure>
210
-
<imgsrc="visualization/visualization_1.png" alt="Visual comparison on RealLQ250">
211
-
<figcaption>Main qualitative comparison on RealLQ250.</figcaption>
212
-
</figure>
213
-
<figure>
214
-
<imgsrc="visualization/visualization_2.png" alt="Additional qualitative comparison on RealLQ250">
215
-
<figcaption>Additional qualitative comparisons on RealLQ250.</figcaption>
216
-
</figure>
217
-
</div>
218
-
219
210
<divclass="table-grid">
220
-
<divclass="table-wrap">
211
+
<divclass="table-wrap compact-table">
221
212
<table>
222
213
<caption>Human-aligned LR-faithfulness evaluation on RealSR.</caption>
223
214
<thead>
@@ -234,7 +225,7 @@ <h2>Experiments</h2>
234
225
</table>
235
226
</div>
236
227
237
-
<divclass="table-wrap">
228
+
<divclass="table-wrap compact-table">
238
229
<table>
239
230
<caption>Ablation study on RealLQ250 using LucidFlux.</caption>
<figcaption>Primary degradation distribution across LucidLR, RealLQ250, RealSR, and DRealSR.</figcaption>
266
-
</figure>
267
-
<figure>
268
-
<imgsrc="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
-
<imgsrc="fig/training_curves_lucidflux_ablation.png" alt="Optimization curves under different reward formulations">
273
-
<figcaption>Optimization curves under different reward formulations on LucidFlux.</figcaption>
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
+
<sectionid="citation" class="paper-citation">
283
322
<h2>Citation</h2>
284
323
<pre><code>@article{fei2026lucidnft,
285
324
title={LucidNFT: LR-Anchored Multi-Reward Preference Optimization for Flow-Based Real-World Super-Resolution},
286
325
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},
288
327
year={2026}
289
328
}</code></pre>
290
329
</section>
291
-
</main>
292
330
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
+
<sectionid="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>
0 commit comments