Object-centric Self-improving Preference Optimization for Text-to-Image Generation (2506.02015v1)

Abstract: Recent advancements in Multimodal LLMs (MLLMs) have significantly improved both image understanding and generation capabilities. Despite these improvements, MLLMs still struggle with fine-grained visual comprehension, particularly in text-to-image generation tasks. While preference optimization methods have been explored to address these limitations in image understanding tasks, their application to image generation remains largely underexplored. To address this gap, we propose an Object-centric Self-improving Preference Optimization (OSPO) framework designed for text-to-image generation by MLLMs. OSPO leverages the intrinsic reasoning abilities of MLLMs without requiring any external datasets or models. OSPO emphasizes the importance of high-quality preference pair data, which is critical for effective preference optimization. To achieve this, it introduces a self-improving mechanism that autonomously constructs object-level contrastive preference pairs through object-centric prompt perturbation, densification and VQA scoring. This process eliminates ambiguous or disproportionate variations commonly found in naively generated preference pairs, thereby enhancing the effectiveness of preference optimization. We validate OSPO on three representative compositional text-to-image benchmarks, demonstrating substantial performance gains over baseline models.