YINYANG-ALIGN: Benchmarking Contradictory Objectives and Proposing Multi-Objective Optimization based DPO for Text-to-Image Alignment
Amitava Das, Yaswanth Narsupalli, Gurpreet Singh, Vinija Jain, Vasu Sharma, Suranjana Trivedy, Aman Chadha, Amit Sheth
2025-02-10
Summary
This paper talks about YinYangAlign, a system designed to evaluate and improve how well AI models align generated images with user prompts while balancing ethical and creative goals.
What's the problem?
Text-to-image (T2I) systems often struggle to create images that perfectly match user instructions while also meeting ethical and artistic standards. Conflicts between goals like being faithful to the prompt and allowing creative freedom can lead to misaligned or inappropriate outputs, as seen in incidents like the Google Gemini controversy.
What's the solution?
The researchers developed YinYangAlign, a framework that measures how well T2I systems balance six conflicting objectives, such as prompt accuracy versus artistic freedom. They also introduced a new optimization method called Contradictory Alignment Optimization (CAO), which improves alignment by managing these trade-offs. The framework includes a detailed dataset with human prompts, aligned and misaligned outputs, and explanations of the contradictions.
Why it matters?
This matters because it helps make AI-generated images more reliable and ethically sound while still being creative. By addressing the challenges of balancing conflicting goals, YinYangAlign can improve the quality and trustworthiness of T2I systems, making them more useful for real-world applications like advertising, design, and content creation.
Abstract
Precise alignment in Text-to-Image (T2I) systems is crucial to ensure that generated visuals not only accurately encapsulate user intents but also conform to stringent ethical and aesthetic benchmarks. Incidents like the Google Gemini fiasco, where misaligned outputs triggered significant public backlash, underscore the critical need for robust alignment mechanisms. In contrast, Large Language Models (LLMs) have achieved notable success in alignment. Building on these advancements, researchers are eager to apply similar alignment techniques, such as Direct Preference Optimization (DPO), to T2I systems to enhance image generation fidelity and reliability. We present YinYangAlign, an advanced benchmarking framework that systematically quantifies the alignment fidelity of T2I systems, addressing six fundamental and inherently contradictory design objectives. Each pair represents fundamental tensions in image generation, such as balancing adherence to user prompts with creative modifications or maintaining diversity alongside visual coherence. YinYangAlign includes detailed axiom datasets featuring human prompts, aligned (chosen) responses, misaligned (rejected) AI-generated outputs, and explanations of the underlying contradictions.