# Ethical-Lens: Curbing malicious usages of open-source text-to-image models

## TL;DR

Ethical-Lens is a framework that helps open-source text-to-image models generate ethical and unbiased images without changing their internal structure.

## Contribution

Introduces Ethical-Lens, a plug-and-play framework for aligning open-source text-to-image models with societal values.

## Key findings

- Ethical-Lens reduces toxicity and bias in generated images while maintaining high image quality.
- The framework outperforms commercial models like DALL·E 3 in ethical alignment and diversity.
- It provides a scalable solution compatible with all open-source text-to-image tools.

## Abstract

The burgeoning landscape of text-to-image models, exemplified by innovations such as Midjourney and DALL·E 3, has revolutionized content creation across diverse sectors. However, these advances bring forth critical ethical concerns, particularly with the misuse of open-source models to generate content that violates societal norms. Addressing this, we introduce Ethical-Lens, a framework designed to facilitate the value-aligned usage of text-to-image tools without necessitating internal model revision. Ethical-Lens ensures value alignment in text-to-image models across toxicity and bias dimensions by refining user commands and rectifying model outputs. Systematic evaluation metrics, combining GPT4-V, HEIM, and FairFace scores, assess alignment capability. Our experiments reveal that Ethical-Lens enhances alignment capabilities to levels comparable with or superior to commercial models such as DALL·E 3, while preserving the quality of generated images. This study indicates the potential of Ethical-Lens to promote the sustainable development of open-source text-to-image tools and their beneficial integration into society.

•Enhances value alignment in text-to-image models without internal modifications•Reduces toxicity and bias in generated images while maintaining high image quality•Outperforms commercial models like DALL·E 3 in ethical alignment and diversity•Provides a plug-and-play framework compatible with all open-source text-to-image tools

Enhances value alignment in text-to-image models without internal modifications

Reduces toxicity and bias in generated images while maintaining high image quality

Outperforms commercial models like DALL·E 3 in ethical alignment and diversity

Provides a plug-and-play framework compatible with all open-source text-to-image tools

Rapid advances in AI text-to-image models have revolutionized content creation but have also raised significant ethical concerns about the generation of harmful or biased content. Commercially available AI models generally have in-built controls that aim to block or reduce the creation of harmful content. Researchers and others, however, may prefer open-source models due to their greater transparency and the ability to build on these models freely and integrate them into other open science projects. A variety of open-source text-to-image models are currently available, but they lack robust ethical controls. Models can be individually tailored to improve their ethical alignment, but this is a costly and laborious approach. Better plug-in-play tools are therefore needed to help open-source model users ensure ethical outputs.

Ethical-Lens introduces a framework to ensure value alignment in text-to-image models, addressing toxicity and bias without altering internal model structures. It outperforms commercial models such as DALL·E 3, offering a scalable solution for open-source tools to generate diverse, unbiased, and high-quality images while adhering to societal norms.