Reproduce “Fully Unsupervised Self-debiasing of Text-to-Image Diffusion Models” 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
Korada Sri Vardhana, Shrikrishna Lolla, Soma Biswas
Researchers from the Indian Institute of Science, Bengaluru, introduce SelfDebias, a fully unsupervised, test-time framework designed to mitigate biases in text-to-image diffusion models without requiring human supervision or model retraining. The framework demonstrates superior fairness performance on demographic and occupation-related biases, and uniquely debiases abstract concepts by automatically discovering semantic modes.
Key insights
- Semantic modes for debiasing can be automatically discovered in a disentangled latent space derived from U-Net activations through hierarchical clustering, removing the need for human annotations or external classifiers.
- Guiding the diffusion process by dynamically minimizing KL divergence to a uniform distribution over these discovered modes effectively mitigates bias at inference time without altering model weights.
- The framework's ability to discover and balance unspecified semantic variations enables debiasing for abstract concepts or non-societal biases where explicit attributes are undefined.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.