What Drives Compositional Generalization in Visual Generative Models?

Static project page for the paper
“What Drives Compositional Generalization in Visual Generative Models?” (Farid et al., 2025).
The live site is a single-page app built with plain HTML/CSS/JS and is intended for GitHub Pages hosting.

Quick Start

• View online: enable GitHub Pages on the main branch (Settings → Pages → Source main / root).
• Preview locally: run a lightweight HTTP server from the repo root and open the reported URL in your browser.

cd gen-comp-gen.github.io
python3 -m http.server 4000
# visit http://127.0.0.1:4000/

Repo Layout

• index.html – main landing page with hero, findings, animated figures, tables, and BibTeX block.
• figures/ – figure assets referenced by the page (PDF/PNG/SVG); see .gitignore for unused variants.
• figs/teaser.jpg – hero image shown in the header and social previews.
• .gitignore – ignores auxiliary figure exports that are not linked from the site.

Editing the Site

1. Update copy or structure directly in index.html.
2. Place new visual assets in figures/ (or figs/ for hero art) and reference them with relative URLs.
3. If you add or remove assets, adjust .gitignore so only unused variants remain ignored.
4. Commit and push; GitHub Pages will redeploy automatically after the push.

Animated figures use a tiny dependency-free script at the bottom of index.html. To add another carousel:

<div id="fig-anim-xyz" class="anim-fig" data-dwell="1500">
  <!-- repeated .anim-step elements -->
</div>
<script>
  mountAnimFigure('fig-anim-xyz');
</script>

Troubleshooting

• Broken images: ensure paths match the filenames (case-sensitive) and that the asset is not listed in .gitignore.
• PDF rendering issues: export to PNG (same basename) if the browser cannot inline the PDF and update the src.
• No GitHub Pages output: verify that Pages is enabled on the correct branch and that the build began in the Pages dashboard.

TODO

High priority

• Add Orbis compositional videos with REAL, MaskGIT, and DiT (one video per novel composition) – Karim
• Document the experimental setup underlying the probe accuracy results and clarify their interpretation on the site.
• Restructure all sections that follow the JEPA section.
• Add a qualitative results section sourced from Shapes3D.
• Add a world models section with quantitative and qualitative results – Karim
• Fix image sizes.
• Proofread the full site copy.
• Move shapes 3d figure to the end of the page.
• Make the shapes 3d figure more readable, clear color legend while scrolling, what should the reader see there.
• Put the videos as the main teaser part.
• Remove the main teaser in the beginning and keep title only.

Less priority

• Add CelebA compostional split images - Karim
• Add CLEVRER videos (111 videos) - Rajat
• Fix language system prompt images - Rajat

For questions, contact the authors at faridk@cs.uni-freiburg.de.