Update index.html

index.html

@@ -48,18 +48,6 @@
   </p>
 </div>
 
-<!-- Interactive Highlight Slider -->
-<div class="container">
-  <h2>Paper Highlights</h2>
-  <div id="highlight-box" style="text-align: center; padding: 30px; border: 1px solid #ddd; border-radius: 10px; background: #fafafa;">
-    <p id="highlight-text" style="font-size: 1.2rem; font-style: italic;"></p>
-  </div>
-  <div class="text-center mt-3">
-    <button onclick="prevHighlight()" class="btn btn-outline-primary">← Prev</button>
-    <button onclick="nextHighlight()" class="btn btn-outline-primary">Next →</button>
-  </div>
-</div>
-
 <!-- Unified Paper Highlights Section -->
 <div class="container" style="max-width: 900px;">
   <h2>Paper Highlights</h2>
@@ -83,7 +71,7 @@
       image: "figures/highlights/table.png"
     },
     {
-      text: "Both modeling paradigms (EnCodec-based latent) show comparable performance with a slight favor toward AR, which also prove to be more robust to the latent representation’s sample rate. FM performance degrade as the number of inference steps decrease. In order to maintain comparable performance with AR, FM requires a large number of inference steps.",
+      text: "Both modeling paradigms (EnCodec-based latent) show comparable performance with a slight favor toward AR, which also proves to be more robust to the latent representation’s sample rate. FM performance degrades as the number of inference steps decreases.",
       image: "figures/highlights/fidelity.png"
     },
     {
@@ -91,15 +79,15 @@
       image: "figures/highlights/control.png"
     },
     {
-      text: "Supervised flow matching is the most robust inpainting method: it yields the smoothest and most coherent edits; zero-shot flow matching is attractive for rapid, prompt-driven edits but needs a small hyper-parameter search per-sample or a better sampling strategy to provide more stable outputs.",
+      text: "Supervised flow matching is the most robust inpainting method: it yields the smoothest and most coherent edits; zero-shot FM is fast but less stable without tuning.",
       image: "figures/highlights/inpainting.png"
     },
     {
-      text: "AR scales better with batch size thanks to KV caching; FM may becomes faster while reducing the number of inference steps, however this comes at the cost of degraded generation quality. Selecting a modeling paradigm therefore hinges on how much quality one is willing to trade for latency.",
+      text: "AR scales better with batch size thanks to KV caching. FM can be faster by reducing inference steps—but this comes at the cost of generation quality.",
       image: "figures/highlights/speed_vs_quality.png"
     },
     {
-      text: "When the number of update steps is capped, FM reaches almost the same FAD, PQ, and CE as in the one-million-step topline using much smaller batches, though its CLAP score keeps improving with scale. The AR model needs a larger token budget per step to match its topline performance and benefits more from large scale training.",
+      text: "When update steps are capped, FM reaches near-topline FAD and PQ even with small batches. AR requires a larger token budget per step to match performance.",
       image: "figures/highlights/training_sensitivity.png"
     }
   ];