150% fairness improvement! New synthetic data method mitigates biases.
Mitigating stereotypical biases in text to image generative systems
Original: Runway Research | Mitigating stereotypical biases in text to image generative systems
Importance: バイアスを軽減する新しい手法の提案は広範な影響を及ぼす可能性があるため。
Summary
State-of-the-art text-to-image models exhibit social biases, overrepresenting lighter-skinned individuals and men. This research proposes a method to mitigate these biases by fine-tuning models on diverse synthetic data varying in skin tones and genders. The diversity finetuned model improves fairness metrics by 150% for perceived skin tone and 97.7% for gender, generating more darker-skinned individuals and women. All prompts and code for generating training images will be released for open research.
Key Points
- Mitigating biases in skin tone and gender
- Fine-tuning using synthetic data
- Fairness metrics improved by 150%
- Increased generation of darker-skinned individuals and women
- Release of prompts and code planned
View developer notes (APIs, breaking changes, migration)
This research employs a fine-tuning method for text-to-image generative models using synthetic data. The synthetic data consists of diverse text prompts combining ethnicities, genders, professions, and age groups. The DFT model significantly improves fairness metrics by 150% for perceived skin tone and 97.7% for gender, resulting in more darker-skinned individuals and women being generated compared to baselines. To enhance research transparency, all text prompts and code for generating training images will be released.
Source: https://runway.com/research/mitigating-stereotypical-biases-in-text-to-image-generative-systems
Outlet: Runway
This article is an AI-generated summary (OpenAI GPT-4o-mini) of publicly available information from Anthropic, OpenAI, Google, Meta, Mistral, DeepSeek, Sakana, and other vendors. The original source URL is always provided in accordance with fair-use citation requirements. Summaries are AI-generated and may contain mistranslations or misinterpretations. Always verify details with the original source.