FAIRT2V: Training-Free Debiasing for Text-to-Video Diffusion Models

TL;DR

This paper addresses gender bias in text-to-video diffusion by showing that encoder-induced associations in prompt embeddings propagate into videos even for neutral prompts. It introduces FairT2V, a training-free debiasing method that steers embeddings along a profession-specific gender axis via anchor-based spherical geodesic transformations, with an adaptive lambda and a dynamic denoising schedule to protect temporal coherence. A video-centric fairness evaluation protocol that combines VideoLLM reasoning with human verification is proposed, and experiments on Open-Sora demonstrate that FairT2V substantially reduces bias with minimal impact on video quality. Together, these contributions advance fairer T2V generation and offer a scalable, inference-time solution with practical implications for deploying bias-aware video synthesis systems.

Abstract

Text-to-video (T2V) diffusion models have achieved rapid progress, yet their demographic biases, particularly gender bias, remain largely unexplored. We present FairT2V, a training-free debiasing framework for text-to-video generation that mitigates encoder-induced bias without finetuning. We first analyze demographic bias in T2V models and show that it primarily originates from pretrained text encoders, which encode implicit gender associations even for neutral prompts. We quantify this effect with a gender-leaning score that correlates with bias in generated videos. Based on this insight, FairT2V mitigates demographic bias by neutralizing prompt embeddings via anchor-based spherical geodesic transformations while preserving semantics. To maintain temporal coherence, we apply debiasing only during early identity-forming steps through a dynamic denoising schedule. We further propose a video-level fairness evaluation protocol combining VideoLLM-based reasoning with human verification. Experiments on the modern T2V model Open-Sora show that FairT2V substantially reduces demographic bias across occupations with minimal impact on video quality.

Figures (10)

• Figure 1: Bias source analysis in text-to-video generation. Neutral prompts are encoded by the text encoder (e.g., CLIP) into embeddings aligned with gender-associated directions, revealing implicit demographic bias in the text-conditioning space.
• Figure 2: Gender-leaning scores (\ref{['eq:score']}) from the CLIP text encoder for 16 occupations, using the prompt sets in \ref{['eq:prompt_sets']}.
• Figure 3: Gender proportions in generated videos for four occupations using neutral prompts (left) and explicit gender cues (right).
• Figure 4: Overview of the FairT2V framework. Neutral prompts are debiased at the embedding level by steering them toward a neutral region between majority- and minority-group anchors. A dynamic denoising schedule applies debiasing only at early diffusion steps to preserve temporal coherence. Fairness is evaluated using a video-level protocol combining LLM-as-a-Judge with human verification.
• Figure 5: Qualitative comparison of text-to-video results for the neutral prompt "A nurse is working." (a) Open-Sora without debiasing, (b) FairDiff Fair-Diffusion, and (c) FairImagen FairImagen are training-free debiasing baselines adapted from text-to-image diffusion models. (d) FairT2V produces more gender-balanced videos while preserving visual quality.