Cross-Modal Bias Transfer in Aligned Video Diffusion Models

Video diffusion models integrate visual, temporal, and textual signals, creating potential pathways for cross-modal bias transfer. This paper studies how alignment tuning affects the transmission of social bias between text and visual modalities in video generation. We evaluate 14,200 text-to-video samples using a cross-modal attribution framework that decomposes bias contributions across input modalities. Quantitative analysis reveals that alignment tuning reduces text-conditioned bias by 24.8%, yet increases visually induced bias carryover by 31.5%, particularly in identity-related scenarios. The results demonstrate that alignment tuning redistributes bias across modalities rather than eliminating it, highlighting the need for modality-aware alignment strategies.