Fairt2v Achieves Training-Free Debiasing of Text-To-Video, Mitigating Gender Bias

Researchers are increasingly aware of demographic biases creeping into artificial intelligence systems, and text-to-video (T2V) generation is no exception. Haonan Zhong, Wei Song, and Tingxu Han, from the University of New South Wales and Nanjing University respectively, alongside Maurice Pagnucco, Jingling Xue, and Yang Song et al., have identified and addressed a significant gender bias within these models , a problem previously under-investigated. Their new framework, FairT2V, uniquely mitigates this bias without requiring any further training of the complex T2V system itself, instead focusing on neutralising biased prompt embeddings originating from the pretrained text encoder. This innovative approach, demonstrated successfully on the Open-Sora model, represents a crucial step towards fairer and more representative video generation, offering a practical solution to a growing ethical concern in AI development.

Their new framework, FairT2V, uniquely mitigates this bias without requiring any further training of the complex T2V system itself, instead focusing on neutralising biased prompt embeddings originating from the pretrained text encoder. This innovative approach, demonstrated successfully on the Open-Sora model, represents a crucial step towards fairer and more representative video generation, offering a practical solution to a growing ethical concern in AI development.

Text-to-video bias stems from pretrained encoders and datasets

Scientists have unveiled FairT2V, a novel training-free framework designed to mitigate demographic biases, specifically gender bias, in text-to-video generation. This groundbreaking study quantifies this effect using a newly developed ‘gender-leaning score’ that directly correlates with bias observed in the generated videos, providing a measurable metric for assessing and addressing the issue. The team achieved substantial reductions in demographic bias across various occupations by strategically applying this debiasing technique only during the early stages of video generation, specifically, the identity-forming steps, through a dynamic denoising schedule. This innovative approach maintains temporal coherence and prevents disruptions to video quality, ensuring visually smooth and consistent results.
To rigorously evaluate the fairness of generated videos, researchers proposed a novel video-level evaluation protocol that combines the reasoning capabilities of VideoLLMs with human verification, addressing the limitations of existing image-based fairness assessments. Experiments conducted on the state-of-the-art Open-Sora text-to-video model revealed that FairT2V significantly reduces demographic bias without compromising video quality. The work establishes a clear link between text encoder biases and their manifestation in generated videos, highlighting the importance of addressing these biases at the source. This research introduces a practical, training-free solution that can be readily integrated into existing text-to-video pipelines, offering a pathway towards more equitable and representative video generation.

Furthermore, the study’s fairness evaluation protocol provides a robust method for assessing demographic bias in videos, moving beyond frame-level analysis to consider temporal dynamics and multi-subject scenes. By leveraging VideoLLMs and human validation, the researchers created a comprehensive system for identifying and mitigating biases, paving the way for more responsible and inclusive AI-generated content. This breakthrough opens exciting possibilities for applications in advertising, education, and entertainment, ensuring that generated videos reflect a more diverse and equitable representation of society.

Gender Bias Quantification in Text-to-Video Systems

Scientists investigated demographic biases within text-to-video (T2V) generation, focusing specifically on gender bias, and developed FairT2V, a training-free debiasing framework. To quantify this, the study pioneered a gender-leaning score, calculated using Equation (5), that demonstrably correlates with bias in the generated videos. This score was derived from analysing 16 common occupations, utilising three prompt sets, gender-neutral, majority-group, and minority-group, defined in Equation (3), to assess the text encoder’s (CLIP) output. The team engineered a method to measure occupation-specific bias using a local bias index, BIoi, calculated as the difference between the dot products of neutral embeddings and those of the majority and minority groups.

A local gender direction, loi, was then determined and normalised, with its sign aligned to the BIoi, providing a vector representing the gender association for each occupation. Subsequently, a global gender axis, goi, was computed by weighting and summing these local directions using Equation (4), effectively establishing a unified representation of gender bias across all occupations. Finally, the gender-leaning score, soi, was calculated by projecting the neutral embedding onto this global axis, revealing the strength and direction of gender association for each occupation. Experiments employed the Open-Sora T2V model to generate 100 videos per occupation using gender-neutral prompts and 200 videos using explicitly gendered prompts, allowing for a direct comparison between embedding-level bias and video output.

Analysis of the generated videos revealed a strong correlation between the calculated gender-leaning scores and the observed gender distributions, confirming that the text-conditioning pathway is a primary source of bias. For instance, occupations with higher male-leaning scores predominantly generated male subjects from neutral prompts, while female-leaning occupations yielded mostly female subjects. To maintain temporal coherence, debiasing is applied only during the early, identity-forming stages of the denoising process via a dynamic schedule.

FairT2V reduces gender bias in video generation

Experiments revealed that occupations predominantly generate male subjects, such as CEO and Doctor, while female-leaning occupations, like Nurse and Teacher, typically yield female subjects. When explicit gender cues are provided, generated identities consistently follow the specified cue across all occupations, aligning with classifier-free guidance which injects the conditioned text embedding at every denoising step. The study demonstrates that the text-conditioning pathway is the primary source of demographic bias in text-to-video diffusion models, motivating the development of a prompt-level debiasing framework. FairT2V mitigates encoder-induced bias by rebalancing gender associations in the text-conditioning space through a novel approach involving anchor-based spherical geodesic transformations.

Researchers constructed two explicit gender anchors by augmenting prompts with majority- and minority-associated attributes, creating embeddings that define an occupation-specific demographic axis while preserving semantics. All text embeddings were l2-normalized prior to angular computations to ensure accurate geodesic transformations. Measurements confirm that the framework obtains a demographically neutral prompt embedding by applying a spherical geodesic transformation between the two gender anchors, inspired by SLERP, moving the embedding along a great-circle trajectory on the unit hypersphere. The angular distance between the anchors, denoted as θ, and a parameter λ determine the spherical position along the demographic axis, enabling both interpolation and controlled extrapolation.

When the anchors are nearly aligned, the framework reduces to a near-identity mapping, while λ ∈[0, 1] performs spherical interpolation; otherwise, it continues along the same geodesic direction. To adaptively determine λ, scientists computed δmaj and δmin, representing the angular proximity of the neutral prompt embedding to each gender anchor, with smaller angles indicating stronger implicit alignment. The adaptive coefficient, λ∗, was then defined to shift the embedding toward the minority anchor along the geodesic direction when closer to the majority anchor, and vice versa. Tests prove that this approach effectively rebalances the anchors along a profession-specific gender axis, preserving prompt meaning as validated by CLIP-based alignment metrics. Furthermore, a dynamic denoising schedule applies debiasing only during early identity-forming steps, maintaining temporal coherence and preventing identity inconsistencies or flicker.

FairT2V mitigates gender bias in text-to-video generation

Scientists have demonstrated a significant advancement in addressing demographic biases within text-to-video (T2V) generation models. Researchers introduced FairT2V, a training-free framework designed to mitigate encoder-induced gender bias without requiring model finetuning. To ensure temporal consistency in generated videos, the framework applies debiasing primarily during the initial, identity-forming stages through a dynamic denoising schedule. Experiments conducted on the Open-Sora T2V model showed that FairT2V substantially reduces demographic bias across various occupations, all while maintaining high video quality.

The significance of these findings lies in the identification of the source of bias and the provision of a practical, training-free method for mitigation. The authors acknowledge that achieving robust debiasing with T5 embeddings remains a challenge, as fine-grained token-level conditioning proves sensitive to perturbations. Future research should focus on extending the approach to address multi-class and cross-demographic biases, broadening the scope of fairness considerations in T2V generation.