YinYang-Align: Benchmarking Contradictory Objectives and Proposing Multi-Objective Optimization based DPO for Text-to-Image Alignment

Author(s)

• Amitava Das, University of South CarolinaFollow
• Yaswanth Narsupalli
• Gurpreet Singh
• Vinija Jain
• Vasu Sharma
• Suranjana Trivedi
• Aman Chadha
• Amit Sheth, University of South CarolinaFollow

ORCID iD

Sheth: https://orcid.org/0000-0002-0021-5293

Document Type

Paper

Subject Area(s)

Generative AI

Abstract

As Text-to-Image (T2I) models become more advanced, they face a fundamental challenge—balancing conflicting alignment goals such as faithfulness vs. artistic freedom, realism vs. stylization, and verifiability vs. creativity. Existing alignment methods often optimize for one objective at the cost of another, leading to inconsistencies in AI-generated images.
/="/">
/="/">In our latest work, YinYang-Align, we introduce a benchmarking framework to systematically evaluate these trade-offs and propose Contradictory Alignment Optimization (CAO)—a multi-objective extension of Direct Preference Optimization (DPO) that enables models to navigate competing alignment goals more effectively.

Digital Object Identifier (DOI)

https://doi.org/10.48550/arXiv.2502.03512

Publication Info

Preprint version arXiv, 2025.

APA Citation

https://doi.org/10.48550/arXiv.2502.03512

Rights

© 2025, The Authors. Licensed under a Creative Commons Attribution 4.0 International License