-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathindex.html
More file actions
439 lines (405 loc) · 22 KB
/
index.html
File metadata and controls
439 lines (405 loc) · 22 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
<!DOCTYPE html>
<html>
<head>
<meta charset="utf-8">
<!-- Meta tags for social media banners, these should be filled in appropriatly as they are your "business card" -->
<!-- Replace the content tag with appropriate information -->
<meta name="description"
content="RobustVisRAG improves Vision-based Retrieval-Augmented Generation (VisRAG) under visual degradations by introducing a causality-guided dual-path encoder that disentangles semantic and degradation factors. Through Non-Causal Distortion Modeling (NCDM) and Causal Semantic Alignment (CSA), it enhances retrieval, generation, and end-to-end robustness without additional inference cost. We further introduce the Distortion-VisRAG benchmark covering synthetic and real-world degradations.">
<meta property="og:title" content="RobustVisRAG: Robust Vision-based Retrieval-Augmented Generation under Visual Degradations" />
<meta property="og:description" content="RobustVisRAG is a causality-aware framework for Vision-based Retrieval-Augmented Generation that disentangles semantic and degradation factors to ensure robust retrieval and generation under visual distortions without additional inference cost." />
<meta property="og:url" content="URL OF THE WEBSITE" />
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X630-->
<meta property="og:image" content="static/image/your_banner_image.png" />
<meta property="og:image:width" content="1200" />
<meta property="og:image:height" content="630" />
<meta name="twitter:title" content="RobustVisRAG: Robust Vision-based Retrieval-Augmented Generation under Visual Degradations">
<meta name="twitter:description" content="RobustVisRAG is a causality-aware framework for Vision-based Retrieval-Augmented Generation that disentangles semantic and degradation factors to ensure robust retrieval and generation under visual distortions without additional inference cost.">
<!-- Path to banner image, should be in the path listed below. Optimal dimenssions are 1200X600-->
<meta name="twitter:image" content="static/images/your_twitter_banner_image.png">
<meta name="twitter:card" content="summary_large_image">
<!-- Keywords for your paper to be indexed by-->
<meta name="keywords" content="Retrieval-Augmented Generation, RAG, Vision-Language Models, VLM, Causal Learning">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>RobustVisRAG: Robust Vision-based Retrieval-Augmented Generation under Visual Degradations
</title>
<link href="https://fonts.googleapis.com/css?family=Google+Sans|Noto+Sans|Castoro" rel="stylesheet">
<link rel="image_src" href="./robustvisrag_files/logo.png">
<link rel="icon"
type="image/x-icon"
href="./robustvisrag_files/logo.png"/>
<link rel="stylesheet" href="static/css/bulma.min.css">
<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">
<link rel="stylesheet" href="static/css/index.css">
<script src="https://ajax.googleapis.com/ajax/libs/jquery/3.5.1/jquery.min.js"></script>
<script src="https://documentcloud.adobe.com/view-sdk/main.js"></script>
<script defer src="static/js/fontawesome.all.min.js"></script>
<script src="static/js/bulma-carousel.min.js"></script>
<script src="static/js/bulma-slider.min.js"></script>
<script src="static/js/index.js"></script>
<script src="https://polyfill.io/v3/polyfill.min.js?features=es6"></script>
<script id="MathJax-script" async
src="https://cdn.jsdelivr.net/npm/mathjax@3/es5/tex-mml-chtml.js">
</script>
</head>
<body>
<!-- Authors -->
<section class="hero">
<div class="hero-body">
<div class="container">
<div class="container has-text-centered">
<h1 class="title is-1 publication-title">
RobustVisRAG: Causality-Aware Vision-Based Retrieval-Augmented Generation under Visual Degradations.
</h1>
<div class="is-size-5 publication-authors">
<div class="author-block"><a href="https://sites.google.com/view/cihsiang/home">I-Hsiang Chen</a> <sup> 1 </sup></div>
<div class="author-block"><a href="https://scholar.google.com/citations?user=oZanP1AAAAAJ&hl=zh-TW"> Yu-Wei Liu</a> <sup> 1</sup></div>
<div class="author-block"><a href=""> Tse-Yu Wu</a> <sup> 1</sup></div>
<div class="author-block"><a href=""> Yu-Chien Chiang</a> <sup> 1</sup></div>
<div class="author-block"><a href="https://scholar.google.com/citations?user=a-i4INAAAAAJ&hl=zh-TW"> Yang Jen-Chieh</a> <sup> 1</sup></div>
<div class="author-block"><a href="https://sites.google.com/view/weitingchen/home"> Wei-Ting Chen</a> <sup> 2 </sup></div>
<div class="is-size-5 publication-authors">
<span class="author-block"><sup>1</sup>National Taiwan University</span>
<span class="author-block"><sup>2</sup>Microsoft</span>
</div>
<!-- <div class="is-size-5 publication-authors">
<span class="author-block"> </sup>* Co-corresponding authors</span>
</div> -->
<div class="is-size-5 publication-authors">
<span class="author-block"> </sup>(CVPR 2026)</span>
</div>
<div class="column has-text-centered">
<div class="publication-links">
<!-- PDF Link. -->
<span class="link-block">
<a href="https://arxiv.org/abs/2602.22013"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="fas fa-file-pdf"></i>
</span>
<span>Paper</span>
</a>
</span>
<span class="link-block">
<a href="https://arxiv.org/abs/2602.22013"
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>
<span class="link-block">
<a href="https://github.com/robustvisrag/RobustVisRAG"
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>
<!-- Dataset Link. -->
<span class="link-block">
<a href="https://hf.co/collections/robustvisrag/distortion-visrag"
class="external-link button is-normal is-rounded is-dark">
<span class="icon">
<i class="far fa-images"></i>
</span>
<span>Data</span>
</a>
</span>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Video -->
<!-- <section class="hero teaser">
<div class="container is-max-desktop">
<div class="hero-body">
<video autoplay muted loop playsinline style="width: 100%; height: auto; border-radius: 12px;">
<source src="./rar_files/Teaser.mp4" type="video/mp4">
Your browser does not support the video tag.
</video>
<h4 class="subtitle has-text-centered">
<span class="methodname">RAR</span> is an iterative image restoration framework that unifies assessment and restoration in a shared latent space.
</h4>
</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">
<h4 class="title is-3">Abstract</h4>
<div class="content has-text-justified">
<p>
Vision-based Retrieval-Augmented Generation (VisRAG) leverages vision-language models (VLMs) to jointly retrieve relevant visual documents and generate grounded answers based on multimodal evidence.
However, existing VisRAG models degrade in performance when visual inputs suffer from distortions such as blur, noise, low light, or shadow, where semantic and degradation factors become entangled within pretrained visual encoders, leading to errors in both retrieval and generation stages.
</p>
<p>
To address this limitation, we introduce RobustVisRAG, a causality-guided dual-path framework that improves VisRAG robustness while preserving efficiency and zero-shot generalization. RobustVisRAG uses a non-causal path to capture degradation signals through unidirectional attention and a causal path to learn purified semantics guided by these signals.
Together with the proposed Non-Causal Distortion Modeling and Causal Semantic Alignment objectives, the framework enforces a clear separation between semantics and degradations, enabling stable retrieval and generation under challenging visual conditions. To evaluate robustness under realistic conditions, we introduce the Distortion-VisRAG dataset, a large-scale benchmark containing both synthetic and real-world degraded documents across seven domains, with 12 synthetic and 5 real distortion types that comprehensively reflect practical visual degradations.
Experimental results show that RobustVisRAG improves retrieval, generation, and end-to-end performance by 7.35%, 6.35%, and 12.40%, respectively, on real-world degradations, while maintaining comparable accuracy on clean inputs.
</p>
</div>
</div>
</div>
</div>
</section>
<!-- Method -->
<section class="section">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h4 class="title is-3 has-text-centered">Proposed Method</h4>
<div class="has-text-centered">
<img style="width: 100%;" src="./robustvisrag_files/Overview.png"
alt="Overview of RobustVisRAG."/>
<div class="content has-text-justified">
<p>
RobustVisRAG enhances Vision-based Retrieval-Augmented Generation (VisRAG) under visual degradations through causality-guided semantic–degradation disentanglement. By explicitly separating degradation and semantic factors inside the vision encoder, our framework suppresses degradation-induced bias while preserving task-relevant representations — without introducing additional inference cost.
</div>
<h3 class="title is-4 has-text-centered">Preliminary</h3>
<div class="columns is-multiline is-variable is-6">
<!-- VisRAG -->
<div class="column is-half">
<p class="content has-text-justified">
<b>Vision-based RAG (VisRAG):</b>
Given a textual query \(q\) (e.g., a question or instruction) and a visual corpus \( \mathcal{V} = \{ X_i \}_{i=1}^{N} \),
VisRAG retrieves the top-\(k\) most relevant document images and generates a response as follows:
\[ \underbrace{R}_{\text{top-}k\ \text{retrieved doc images}}
= \mathcal{R}\!\big(q,\; \mathcal{E}_r(\mathcal{V})\big)
\qquad
Y = \mathcal{G}\!\big(q,\; \mathcal{E}_g(R)\big)
\]
where \( X_i \) denotes the \(i\)-th document image in the corpus,
\( \mathcal{E}_r \) and \( \mathcal{E}_g \) represent the retrieval and generation encoders,
and \( \mathcal{R}(\cdot) \) and \( \mathcal{G}(\cdot) \) denote the retrieval and generation modules, respectively.
Under visual degradations, corrupted representations lead to unstable retrieval and generation.
</p>
</div>
<!-- SCM -->
<div class="column is-half">
<p class="content has-text-justified">
<b>Structural Causal Model (SCM):</b>
We model semantic content \( S \) and degradation \( D \) as independent causes of the observed image \( X \).
In standard encoders, both factors are entangled after conditioning on \( X \), leading to degradation leakage in downstream predictions \( A \).
The overall causal structure can be summarized as:
\[
S \rightarrow X \leftarrow D, \qquad
X \rightarrow Z \rightarrow A
\]
Our goal is to block the <span class="tag is-light" style="background-color: #ffeef0; color: #b30033;">non-causal path</span>:
\[
D \rightarrow X \rightarrow Z \rightarrow A,
\]
while preserving the <span class="tag is-light" style="background-color: #e7f6e7; color: #1e5e1e;">causal path</span>:
\[
S \rightarrow X \rightarrow Z \rightarrow A
\]
</p>
</div>
</div>
<h3 class="title is-4 has-text-centered">RobustVisRAG</h3>
<div class="columns is-multiline is-variable is-6">
<!-- Non-Causal Path -->
<div class="column is-half">
<p class="content has-text-justified">
<span class="tag is-light" style="background-color: #ffeef0; color: #b30033;"><b>Non-Causal Path</b></span>
We introduce a dedicated non-causal token to aggregate degradation signals via <i>unidirectional attention</i>, producing a degradation representation: \( Z_{zeg} \)
Patch tokens do not attend back to this token, preventing degradation leakage into semantic representations.
</p>
<!-- <hr class="dashed-line"> -->
<p class="content has-text-justified">
<b>Non-Causal Distortion Modeling (NCDM):</b> To structure the degradation subspace, we apply a triplet contrastive objective:
\[
\mathcal{L}_{\text{NCDM}}
= \max\big(0,\;
\|Z^a_{\text{deg}} - Z^p_{\text{deg}}\|_2^2
- \|Z^a_{\text{deg}} - Z^n_{\text{deg}}\|_2^2
+ \delta \big),
\]
This enforces clustering of identical distortion types while separating different degradations.
</p>
</div>
<!-- Causal Path -->
<div class="column is-half">
<p class="content has-text-justified">
<span class="tag is-light" style="background-color: #e7f6e7; color: #1e5e1e;"><b>Causal Path</b></span>
The causal branch aggregates patch tokens bidirectionally to produce purified semantic embeddings: \( Z_{sem} \)
This path is isolated from degradation tokens and is the only representation used at inference.
</p>
<!-- <hr class="dashed-line"> -->
<p class="content has-text-justified">
<b>Causal Semantic Alignment (CSA):</b> To ensure degradation-invariant semantics, we aligns degraded semantic embeddings with their clean counterparts while enforcing independence between semantic and degradation representations.:
\[
\begin{aligned}
\mathcal{L}_{\text{CSA}}
&= \frac{1}{T} \sum_{i=1}^{T}
\Big[(1 - \langle Z_{\text{sem},i}^{\text{deg}}, Z_{\text{sem},i}^{\text{clean}} \rangle)
+ \big|\langle Z_{\text{sem},i}^{\text{deg}}, Z_{\text{deg}}^{\text{deg}} \rangle\big|\Big] \\
&\quad + \frac{1}{T} \sum_{i=1}^{T}
\big\| Z_{\text{sem},i}^{\text{deg}} - Z_{\text{sem},i}^{\text{clean}} \big\|_2^2.
\end{aligned}
\]
CSA enforces semantic consistency while discouraging degradation contamination.
</p>
</div>
</div>
</div>
</section>
<!-- Dataset -->
<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">
<h4 class="title is-3">Distortion-VisRAG Dataset</h4>
<div class="has-text-centered">
<img style="width: 100%;" src="./robustvisrag_files/DVisRAG.png" />
<div class="content has-text-justified">
<p>
Distortion-VisRAG (DVisRAG) extends VisRAG with large-scale synthetic and real-world degradations
to systematically evaluate robustness in vision-based RAG systems.
The benchmark contains 367,608 Q–D pairs across seven document domains,
including 12 synthetic distortion types (five severity levels) and 5 real-world recapture conditions.
</p>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Ret -->
<section class="section">
<div class="hero-body">
<div class="container is-max-desktop">
<h4 class="title is-3 has-text-centered">Retrieval Results</h4>
<div class="content has-text-justified">
<p>
RobustVisRAG consistently improves retrieval accuracy under both synthetic and real-world degradations, demonstrating strong distortion-invariant semantic representation.
</p>
</div>
<div id="results-carousel" class="carousel results-carousel">
<!-- Sample 1 -->
<div class="item">
<div>
<img src="robustvisrag_files/ret/ret_sample1_.png"/>
</div>
</div>
<!-- Sample 2 -->
<div class="item">
<div>
<img src="robustvisrag_files/ret/ret_sample2_.png"/>
</div>
</div>
<!-- Sample 3 -->
<div class="item">
<div>
<img src="robustvisrag_files/ret/ret_sample3_.png"/>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Gen -->
<section class="hero is-light">
<div class="hero-body">
<div class="container is-max-desktop">
<h4 class="title is-3 has-text-centered">Generation Results</h4>
<div class="content has-text-justified">
<p>
RobustVisRAG significantly improves answer accuracy under visual degradations by preserving clean semantic representations for generation.
</p>
</div>
<div id="results-carousel" class="carousel results-carousel">
<!-- Sample 1 -->
<div class="item">
<div>
<img src="robustvisrag_files/gen/gen_sample1_.png"/>
</div>
</div>
<!-- Sample 2 -->
<div class="item">
<div>
<img src="robustvisrag_files/gen/gen_sample2_.png"/>
</div>
</div>
<!-- Sample 3 -->
<div class="item">
<div>
<img src="robustvisrag_files/gen/gen_sample3_.png"/>
</div>
</div>
</div>
</div>
</div>
</section>
<!-- Qualitative Results -->
<section class="section">
<div class="container is-max-desktop">
<div class="content has-text-justified">
<h2 class="title is-3 has-text-centered">Quantitative Results</h2>
<p>
We evaluate RobustVisRAG across retrieval, generation, and end-to-end settings
under clean, synthetic, and real-world degradations.
Our method consistently improves robustness without additional inference cost.
</p>
<div class="columns is-variable is-6">
<!-- Left column -->
<div class="column is-half">
<div class="image-with-caption mb-5">
<div class="caption"><b>Overall retrieval performance (MRR@10).</b></div>
<img style="width: 100%;" src="./robustvisrag_files/Exp_Ret.png"/>
</div>
<div class="image-with-caption">
<div class="caption"><b>End-to-end retrieval–generation performance.</b></div>
<img style="width: 100%;" src="./robustvisrag_files/Exp_E2E.png"/>
</div>
</div>
<!-- Right column -->
<div class="column is-half">
<div class="image-with-caption">
<div class="caption"><b>Overall generation performance (Accuracy).</b></div>
<img style="width: 100%;" src="./robustvisrag_files/Exp_Gen.png"/>
</div>
</div>
</div>
</div>
</div>
</section>
<section class="section" id="BibTeX">
<div class="container content is-max-desktop">
<h4 class="title is-4 has-text-centered">BibTeX</h4>
<pre><code>@misc{chen2026robustvisragcausalityawarevisionbasedretrievalaugmented,
title={RobustVisRAG: Causality-Aware Vision-Based Retrieval-Augmented Generation under Visual Degradations},
author={I-Hsiang Chen and Yu-Wei Liu and Tse-Yu Wu and Yu-Chien Chiang and Jen-Chien Yang and Wei-Ting Chen},
year={2026},
eprint={2602.22013},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2602.22013},
}
</code></pre>
</div>
</section>
<!-- Footer -->
<footer class="footer">
<div class="container">
<div class="content has-text-centered">
<a class="icon-link" href="https://arxiv.org/abs/2602.22013">
<i class="fas fa-file-pdf"></i>
</a>
<a class="icon-link" href="https://github.com/robustvisrag/RobustVisRAG" class="external-link" disabled>
<i class="fab fa-github"></i>
</a>
</div>
</div>
</footer>
</body>
</html>