DebiasPI: Inference-time Debiasing by Prompt Iteration of a Text-to-Image Generative Model

TL;DR

DebiasPI introduces an inference-time prompt-iteration framework to actively control demographic attribute distributions in text-to-image generation without retraining. It tracks attribute occurrences via internal model beliefs or external classifiers and iteratively depletes target attribute bins to converge toward a user-specified distribution $Q$, using a subgrouping strategy to speed convergence. Empirical results show DebiasPI can achieve balanced race or gender representations, but skin-tone coverage remains challenging and multi-attribute balancing can incur trade-offs; ablations illustrate the necessity of tracking choices to avoid biased outputs. The work provides a codebook for annotation, evaluation tools, and datasets, and releases code to facilitate benchmarking in future research.

Abstract

Ethical intervention prompting has emerged as a tool to counter demographic biases of text-to-image generative AI models. Existing solutions either require to retrain the model or struggle to generate images that reflect desired distributions on gender and race. We propose an inference-time process called DebiasPI for Debiasing-by-Prompt-Iteration that provides prompt intervention by enabling the user to control the distributions of individuals' demographic attributes in image generation. DebiasPI keeps track of which attributes have been generated either by probing the internal state of the model or by using external attribute classifiers. Its control loop guides the text-to-image model to select not yet sufficiently represented attributes, With DebiasPI, we were able to create images with equal representations of race and gender that visualize challenging concepts of news headlines. We also experimented with the attributes age, body type, profession, and skin tone, and measured how attributes change when our intervention prompt targets the distribution of an unrelated attribute type. We found, for example, if the text-to-image model is asked to balance racial representation, gender representation improves but the skin tone becomes less diverse. Attempts to cover a wide range of skin colors with various intervention prompts showed that the model struggles to generate the palest skin tones. We conducted various ablation studies, in which we removed DebiasPI's attribute control, that reveal the model's propensity to generate young, male characters. It sometimes visualized career success by generating two-panel images with a pre-success dark-skinned person becoming light-skinned with success, or switching gender from pre-success female to post-success male, thus further motivating ethical intervention prompting with DebiasPI.

Figures (8)

• Figure 1: AI model visualizing the news headline: "From School Janitor to Esteemed School Superintendent" without (left) and with (middle, right) prompt intervention. The two-panel image on the left is supposed to show the same person at different stages of their life, but the janitor is depicted as Black and the superintendent as White.
• Figure 2: Three levels of prompting: The Baseline Prompt does not include any ethical intervention. The Prompt with Attribute List mentions attribute choices, while the Prompt with Attribute Distribution asks the model to choose attributes according to a desired distribution. A use case could be for the AI model to generate as many female as male entrepreneur pictures.
• Figure 3: Overview of the proposed Debiasing by Prompt Iteration (DebiasPI) process.
• Figure 4: ChatGPT prompt for creation of success story news headlines.
• Figure 5: Two-panel generated images obtained with attribute list prompts. The panel showing career success often showed a lighter-skinned person, usually a male.