Creating Fair and Inclusive AI-Generated Images with ITI-GEN1

By Dr Silpaja Chandrasekar, PhDReviewed by Susha Cheriyedath, M.Sc.Sep 14 2023

In an article recently submitted to the ArXiv* server, researchers addressed biases in text-to-image generative models. They introduced inclusive models that ensured balanced representations of attributes in generated images. The Inclusive Text-to-Image Generation-1 (ITI-GEN1) was an innovative approach that used prompt embeddings and reference images without model fine-tuning. ITI-GEN1 significantly improved upon existing models for generating inclusive images.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.

Background

In recent years, advancements in generative modeling and access to multimodal datasets have enabled text-based visual content creation. However, existing text-to-image models often inherit biases from training data and lack inclusiveness. To tackle this, researchers are exploring innovative methods like ITI-GEN, which leverage reference images and prompt embeddings to achieve inclusiveness without extensive model retraining or complex prompt specification.

Previous research has extensively explored text-based image generation using various model architectures and datasets. Diffusion-based models have gained attention for their success in handling large multimodal datasets. However, these models often inherit biases from their training data that raise questions about inclusiveness in generative models. While fairness in discriminative models has been well-studied, fair generative models are relatively limited. Some attempts to address bias in generative models have involved Generative Adversarial Network (GAN) based approaches and hard prompt searching methods, but they have limitations.

Proposed Method

ITI-GEN emerges as a groundbreaking approach to create inclusive prompts that capture a wide range of attributes and their combinations in the pursuit of Inclusive Text-to-Image Generation. This becomes especially valuable when dealing with attributes that are challenging to describe using conventional language or are underrepresented. ITI-GEN employs reference images as guiding beacons to offer unambiguous specifications of diverse attributes. The framework of ITI-GEN is structured in three main sections: an overview, a detailed discussion of the learning strategy, and a focus on the essential properties of the approach.

At its core, ITI-GEN aims to address the challenge of producing equal or controllable numbers of images representing various attribute combinations. This is achieved by introducing learnable inclusive tokens injected into the original prompts. These inclusive tokens serve as key components to represent specific attribute categories. ITI-GEN optimizes prompts entirely within the continuous embedding space, allowing for a more flexible and inclusive approach rather than relying on explicit language descriptions. It leverages reference images as a valuable resource to guide prompt learning and ensure that the prompts align with the attributes of these images. The result is a robust and adaptable framework that fosters inclusiveness across many attributes and offers fine control over generated image distributions.

ITI-GEN utilizes direction alignment and semantic consistency losses to guide the prompts effectively in the learning process. Direction alignment aligns the prompts' directions with those of the reference images by facilitating the learning of nuanced differences between various attribute categories. To address language drift and ensure that the prompts maintain their linguistic integrity, a semantic consistency loss is introduced. The optimization strategy involves pair-wise updates to the embeddings of inclusive tokens for different attribute categories to provide a comprehensive solution for prompt learning. Additionally, ITI-GEN demonstrates remarkable generalizability across different models and offers exceptional efficiency by making it a versatile and practical tool for inclusive text-to-image generation.

Experimental Analysis

The experimental analysis showcases the compatibility of ITI-GEN with various state-of-the-art models and techniques by demonstrating its ability to promote inclusiveness and attribute control in image generation. The versatility of ITI-GEN in accommodating different models and conditions is highlighted, enhancing their capabilities without major modifications. The illustration of compatibility with ControlNet, a model capable of conditioning various inputs beyond text, highlights the versatility of ITI-GEN.

It is to be noted that by employing inclusive tokens designed for specific attributes, such as skin tone, ITI-GEN expands ControlNet's capabilities to generate images that manifest the desired attribute while maintaining distributional control. Furthermore, it is highlighted that ITI-GEN can also be integrated with a method for image editing guided by textual instructions called as InstructPix2Pix (IP2P). By utilizing ITI-GEN's attribute-specific tokens, the demonstration illustrates how it can enhance IP2P's inclusiveness on the target attribute, ensuring minimal interference with other image features like clothing and background.

The compatibility and synergy between ITI-GEN and these advanced models and techniques enable various applications. These include fine-grained attribute control and enhanced inclusiveness, achieved with minimal additional complexity or changes to the original models. This flexibility makes ITI-GEN a valuable tool for addressing various challenges in image generation and promoting the generation of diverse, inclusive, and controlled images.

Conclusion

In summary, ITI-GEN introduces a novel method for inclusive text-to-image generation that leverages reference images to enhance inclusiveness. ITI-GEN is a versatile and efficient approach that scales to multiple attributes and domains, supports complex prompts, and is compatible with existing text-to-image generative models. Extensive experiments showcase its effectiveness across various attributes. However, some limitations remain, including challenges with subtle attributes and the need for reference images. Mitigation strategies could involve integrating ITI-GEN with models offering robust controls.

*Important notice: arXiv publishes preliminary scientific reports that are not peer-reviewed and, therefore, should not be regarded as conclusive, guide clinical practice/health-related behavior, or treated as established information.