AI Editorial Model Training

This research project focuses on improving the fairness, diversity, and quality of AI-generated images by training models to produce unbiased representations of individuals across various professions. The study explores challenges in generating fair and balanced depictions of workers, particularly in roles such as construction, fast food, and social work, where biases in gender representation and artistic styles were prevalent. Key objectives include refining captioning strategies to reduce bias, enhancing the diversity of training data, and addressing technical limitations that influence image generation quality. The research employs tools like IP Adapter to mask or modify biased features in generated images, while also leveraging negative prompting to guide the model toward more cohesive and unbiased outputs. The findings highlight significant progress in mitigating biases but also reveal ongoing challenges, such as the subjective nature of defining "unbiased" outcomes and the persistent influence of artistic styles and language prompts on image generation. This research underscores the importance of iterative testing, diverse datasets, and advanced tools to develop AI models capable of producing fair and inclusive representations across professions. This project serves as a foundation for future work aimed at improving AI's ability to generate equitable and realistic depictions of individuals in various roles, ultimately contributing to more inclusive and unbiased applications of AI technology.

Faculty and Staff Leads

Zach Wise

Professor, Journalism

Emmy winning interactive producer & Associate Professor @NorthwesternU, @KnightLab. Formerly of The New York Times. Creator of TimelineJS & StoryMapJS

Project Details

2024 Fall
AI Editorial Model Training

Description

Tools like Stable Diffusion and Midjourney are shaking up the art world in large part because they are accessible for almost anyone to use. The limits of what they can create is still being tested. Many journalism articles lack visuals which are essential in today's platform centric media landscape. We have several projects that utilize AI art for journalism but the trained model is problematic. This project builds on previous quarter's work of training an editorial model that minimizes stereotypes, sexism and racism for editorial use.

Important Questions
  • Can a written article be turned into a descriptive prompt for AI art?
  • What are the implications for concept art generated by AI running with a reported article?
  • What are the limitations of content and concept for generating news art?
Outcome

By the end of the quarter, students will have trained LORA models that change the output of a Stable Diffusion model to be less problematic

2025 Winter
AI Editorial Model Training

Description

Tools like Stable Diffusion and Midjourney are shaking up the art world in large part because they are accessible for almost anyone to use. The limits of what they can create is still being tested. Many journalism articles lack visuals which are essential in today's platform centric media landscape. We have several projects that utilize AI art for journalism but the trained model is problematic. This project builds on previous quarter's work of training an editorial model that minimizes stereotypes, sexism and racism for editorial use.

Important Questions
  • How can we scale training to include more types of situations effeciently
  • What is the best way of diceminating the results of our training and communicate the problem to a larger population
  • What are the limitations of content and concept for generating news art?
Outcome

By the end of the quarter, students will have trained LORA models that change the output of a Stable Diffusion model to be less problematic

2025 Spring
AI Editorial Model Training

Description

This research project focuses on improving the fairness, diversity, and quality of AI-generated images by training models to produce unbiased representations of individuals across various professions. The study explores challenges in generating fair and balanced depictions of workers, particularly in roles such as construction, fast food, and social work, where biases in gender representation and artistic styles were prevalent. Key objectives include refining captioning strategies to reduce bias, enhancing the diversity of training data, and addressing technical limitations that influence image generation quality. The research employs tools like IP Adapter to mask or modify biased features in generated images, while also leveraging negative prompting to guide the model toward more cohesive and unbiased outputs. The findings highlight significant progress in mitigating biases but also reveal ongoing challenges, such as the subjective nature of defining "unbiased" outcomes and the persistent influence of artistic styles and language prompts on image generation. This research underscores the importance of iterative testing, diverse datasets, and advanced tools to develop AI models capable of producing fair and inclusive representations across professions. This project serves as a foundation for future work aimed at improving AI's ability to generate equitable and realistic depictions of individuals in various roles, ultimately contributing to more inclusive and unbiased applications of AI technology.

Important Questions
  • How can we scale training to include more types of situations effeciently
  • What is the best way of diceminating the results of our training and communicate the problem to a larger population
Outcome

By the end of the quarter, students will have trained LORA models that change the output of a Stable Diffusion model to be less problematic