SensIMG-Bench/index.html at main · danielle34/SensIMG-Bench · GitHub

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
<!DOCTYPE html>
<html lang="en">
  <head>
    <meta charset="utf-8" />
    <meta name="viewport" content="width=device-width, initial-scale=1" />
    <meta
      name="description"
      content="SensIMG-Bench is a literature-grounded benchmark for evaluating sensory appropriateness in AI-generated images for children with autism spectrum disorder."
    />
    <title>SensIMG-Bench | Project Page</title>
    <link rel="stylesheet" href="styles.css" />
  </head>
  <body>
    <header class="site-header">
      <nav class="topbar" aria-label="Project links">
        <a href="#abstract">Abstract</a>
        <a href="#overview">Overview</a>
        <a href="#checks">Checks</a>
        <a href="#examples">Examples</a>
        <a href="#method">Method</a>
        <a href="#citation">Citation</a>
      </nav>
    </header>

    <main>
      <section class="paper-header reveal">
        <h1>
          SensIMG-Bench: A Literature-Grounded Benchmark for
          Sensory-Appropriate AI-Generated Images
        </h1>
        <p class="authors">
          Morayo Danielle<sup>1</sup>,
          Paul Pu Liang<sup>2</sup>
        </p>
        <p class="affiliations">
          <sup>1</sup>Howard University<br />
          <sup>2</sup>MIT Media Lab, Massachusetts Institute of Technology
        </p>
        <div class="actions" aria-label="Primary project links">
          <a class="button primary" href="paper/SensIMG_Bench_paper.pdf" target="_blank" rel="noopener" aria-label="Open paper PDF">
            <svg aria-hidden="true" viewBox="0 0 24 24">
              <path d="M6 2h8l4 4v16H6z" fill="none" stroke="currentColor" stroke-width="1.8" />
              <path d="M14 2v5h5" fill="none" stroke="currentColor" stroke-width="1.8" />
              <path d="M8.5 12h7M8.5 16h7M8.5 8.5h3" fill="none" stroke="currentColor" stroke-width="1.8" stroke-linecap="round" />
            </svg>
            Paper
          </a>
          <a class="button" href="https://github.com/danielle34/SensIMG-Bench" target="_blank" rel="noopener" aria-label="Open code on GitHub">
            <svg aria-hidden="true" viewBox="0 0 24 24">
              <path fill="currentColor" d="M12 .5A11.5 11.5 0 0 0 8.36 22.9c.58.11.79-.25.79-.56v-2.02c-3.22.7-3.9-1.38-3.9-1.38-.53-1.34-1.29-1.7-1.29-1.7-1.05-.72.08-.71.08-.71 1.16.08 1.77 1.19 1.77 1.19 1.04 1.76 2.72 1.25 3.38.96.11-.75.41-1.25.74-1.54-2.57-.29-5.27-1.28-5.27-5.72 0-1.26.45-2.3 1.19-3.11-.12-.29-.52-1.47.11-3.07 0 0 .98-.31 3.18 1.19A11.05 11.05 0 0 1 12 6.24c.98 0 1.96.13 2.88.39 2.2-1.5 3.17-1.19 3.17-1.19.64 1.6.24 2.78.12 3.07.74.81 1.19 1.85 1.19 3.11 0 4.45-2.71 5.43-5.29 5.72.42.36.79 1.07.79 2.16v3.2c0 .31.21.67.8.56A11.5 11.5 0 0 0 12 .5Z" />
            </svg>
            Code
          </a>
          <a class="button" href="#" aria-label="Dataset Hugging Face placeholder link">Dataset</a>
        </div>

        <figure id="overview" class="overview-figure">
          <img
            src="site_assets/overview.png"
            alt="Overview diagram of the SensIMG-Bench data generation, visual feature scoring, and LLM synthesis pipeline"
          />
          <figcaption>
            Figure 1. SensIMG-Bench converts literature-derived visual
            sensitivity criteria into feature-level measurements, then compares
            benchmark scores with LLM-based suitability judgments.
          </figcaption>
        </figure>

        <div class="study-strip" aria-label="Study design summary">
          <div>
            <strong>15,000 images</strong>
            <span>Generated across five open-weights text-to-image models.</span>
          </div>
          <div>
            <strong>1,000 prompts</strong>
            <span>Derived from ASD-focused generative visual systems.</span>
          </div>
          <div>
            <strong>60,000 Qwen calls</strong>
            <span>Four VLM prompting strategies evaluated on the same corpus.</span>
          </div>
        </div>
      </section>

      <section id="abstract" class="text-section reveal">
        <h2>Abstract</h2>
        <p>
          Over 96% of children with autism spectrum disorder experience sensory
          sensitivities, yet AI-generated images used in educational and
          therapeutic tools are often produced without measuring sensory safety.
          SensIMG-Bench translates 50 peer-reviewed studies of visual
          sensitivity in ASD into six pixel-level features: spatial information,
          brightness, color saturation, spatial frequency, color entropy, and
          pattern density.
        </p>
        <p>
          We evaluate SDXL, FLUX.1-dev, Stable Diffusion 1.5, Stable Diffusion
          3.5 Large, and PixArt-Sigma on 1,000 prompts across eight content
          categories and three audience-specification conditions, producing
          15,000 generated images. We also evaluate Qwen2.5-VL-7B on the same
          corpus across four prompting strategies, finding that explicit
          feature grounding improves alignment with benchmark decisions compared
          with standard zero-shot prompting.
        </p>
      </section>

      <section id="checks" class="section reveal">
        <h2>What the Benchmark Checks</h2>
        <p class="section-intro">
          For each generated image, SensIMG-Bench computes interpretable visual
          attributes grounded in prior work on sensory sensitivity. These checks
          make the evaluation more explicit than asking a model to generate a
          suitable image in a single zero-shot step.
        </p>
        <p class="section-intro">
          The benchmark is designed to answer three linked evaluation problems:
          whether open-weights image generators stay within the literature-based
          sensory envelope, whether audience-specific prompting moves images
          toward that envelope, and whether Qwen2.5-VL can recover the same
          clinical feature signal from the image alone or requires explicit
          pixel-level grounding.
        </p>

        <div class="checks-grid">
          <article>
            <h3>Brightness</h3>
            <p>Mean luminance and high-intensity regions.</p>
          </article>
          <article>
            <h3>Color Saturation</h3>
            <p>Overall color intensity in generated images.</p>
          </article>
          <article>
            <h3>Color Entropy</h3>
            <p>Hue distribution and chromatic disorganization.</p>
          </article>
          <article>
            <h3>Spatial Information</h3>
            <p>Edge density, gradients, and visual clutter.</p>
          </article>
          <article>
            <h3>Pattern Density</h3>
            <p>Repeating structure in the 2-5 cpd band.</p>
          </article>
          <article>
            <h3>Spatial Frequency</h3>
            <p>Fine-detail energy in the 4-8 cpd band.</p>
          </article>
        </div>
      </section>

      <section id="examples" class="section reveal">
        <h2>Interactive Feature Heatmaps</h2>
        <p class="section-intro">
          The paper visualizes the benchmark by applying each feature check to
          representative generated images. The grid below is generated from the
          same source images and feature functions used for the paper figure.
          Hover over a panel to inspect it, or click to expand the cell.
        </p>

        <figure class="interactive-heatmap">
          <div id="heatmap-grid" class="heatmap-grid" aria-label="Interactive SensIMG-Bench feature heatmap grid"></div>
          <figcaption>
            Rows show one representative image per generator. Columns show the
            original image and six SensIMG-Bench feature heatmaps. A check mark
            indicates the feature passed; a cross indicates that the feature was
            flagged at the benchmark threshold.
          </figcaption>
        </figure>
      </section>

      <section id="method" class="text-section reveal">
        <h2>Approach</h2>
        <p>
          SensIMG-Bench is organized around a compact pipeline: literature
          findings define image-level checks, controlled prompts generate an
          image corpus, and the benchmark computes pass/fail-style sensory-risk
          decisions with feature-level flags. The goal is not to replace human
          review, but to provide a structured first pass for finding generated
          images that may be visually unsuitable.
        </p>
        <p>
          A central finding motivating the benchmark is that generic zero-shot
          image generation does not reliably satisfy sensory constraints. The
          benchmark makes those constraints explicit and checks them consistently
          across images.
        </p>
      </section>
    </main>

    <footer id="citation" class="footer reveal">
      <div class="footer-grid">
        <div>
          <h2>Citation</h2>
          <pre><code>@misc{adeyemi2026sensimgbench,
  title  = {SensIMG-Bench: A Literature-Grounded Benchmark for Sensory-Appropriate AI-Generated Images},
  author = {Adeyemi, Morayo Danielle and Liang, Paul Pu},
  year   = {2026},
  note   = {Public release}
}</code></pre>
        </div>
        <div class="contact">
          <h2>Contact</h2>
          <p>Morayo Danielle</p>
          <a href="mailto:morayoadeyemi2c@gmail.com">morayoadeyemi2c@gmail.com</a>
        </div>
      </div>
    </footer>

    <div id="heatmap-modal" class="heatmap-modal" aria-hidden="true">
      <button class="modal-backdrop" type="button" aria-label="Close expanded heatmap"></button>
      <div class="modal-panel" role="dialog" aria-modal="true" aria-labelledby="modal-title">
        <button class="modal-close" type="button" aria-label="Close expanded heatmap">×</button>
        <img id="modal-image" src="" alt="" />
        <div class="modal-copy">
          <p id="modal-model" class="modal-model"></p>
          <h3 id="modal-title"></h3>
          <p id="modal-description"></p>
        </div>
      </div>
    </div>

    <script src="script.js"></script>
  </body>
</html>