Reproduce “FAIRT2V: Training-Free Debiasing for Text-to-Video 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
Haonan Zhong, Wei Song, Tingxu Han, Maurice Pagnucco, Jingling Xue, Yang Song
FAIRT2V introduces a training-free framework that mitigates demographic bias in text-to-video diffusion models by applying a spherical geodesic transformation to prompt embeddings. The method significantly reduces bias across occupations while preserving video quality and temporal coherence, outperforming existing debiasing techniques.
Key insights
- Implicit demographic bias in text-to-video models primarily stems from the text encoder's conditioning pathway, not the video generation process itself.
- Bias can be effectively mitigated in a training-free manner by precisely manipulating prompt embeddings via spherical geodesic transformations based on demographic anchors.
- Maintaining video quality and temporal coherence during debiasing requires a dynamic approach, selectively applying debiased embeddings only during critical initial diffusion steps.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.