A 150% improvement in biases related to skin tone and gender! A new method for fairness using synthetic data.
Runway Research | Mitigating stereotypical biases in text to image generative systems
Original: Runway Research | Mitigating stereotypical biases in text to image generative systems
Importance: 生成モデルの偏見軽減に向けた重要な研究成果
Summary
State-of-the-art text-to-image models show social biases, over-representing lighter skin tones and men. This study proposes a method to mitigate such biases by fine-tuning models on synthetic data with diverse skin tones and genders. The diversity fine-tuned (DFT) model improves group fairness metrics by 150% for perceived skin tone and 97.7% for gender. All text prompts and code for generating training images will be released.
Key Points
- Method to mitigate biases in skin tone and gender
- Fine-tuning models using synthetic data
- Significant improvements in fairness metrics
- Code and prompts for training will be released
View developer notes (APIs, breaking changes, migration)
This study proposes a method for mitigating biases in text-to-image models by fine-tuning on diverse synthetic data regarding skin tones and genders. Text prompts are generated from combinations of ethnicities, genders, professions, and age groups, resulting in diverse synthetic datasets. The diversity fine-tuned (DFT) model demonstrates a 150% improvement in fairness metrics for skin tone and a 97.7% improvement for gender. All text prompts and code for generating training images will be released shortly.
Source: https://runwayml.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.