Sina Malakouti

Sina Malakouti

he/him

CS PhD Candidate @ UPitt

Role Bias in Diffusion Models: Diagnosing and Mitigating through Intermediate Decomposition

Published in The Thirty-Ninth Annual Conference on Neural Information Processing Systems (NeurIPS), 2025

This work introduces RoleBench, a benchmark for evaluating compositional generalization in text-to-image models through action-based relations. It identifies a systematic failure—RoleCollapse—where models default to frequent reversed relations, and shows that a lightweight intermediate fine-tuning approach (ReBind) can significantly reduce role bias and improve compositional generation.