Reproduce “Diverse Diffusion: Enhancing Image Diversity in Text-to-Image Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Mariia Zameshina, Olivier Teytaud, Laurent Najman
Diverse Diffusion increases the variety of images produced for a fixed text prompt by selecting mutually separated latent vectors before Stable Diffusion generation. Without retraining or demographic labels, its pooling-based selection improves several color, perceptual, and demographic-coverage measures, although it does not guarantee fairness, semantic diversity, or stronger prompt adherence.
Key insights
- Latent-space separation can influence broad image properties such as dominant color and demographic appearance, even though the selection objective does not explicitly encode either attribute.
- The benefits are clearest when baseline diversity is low or batch sizes are larger; pooling-based methods generally outperform ordinary sampling in these settings.
- Greater latent separation is not equivalent to guaranteed semantic diversity: generated images may differ in irrelevant ways, fail to satisfy the prompt, or shift demographic composition inconsistently.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.