Core Viewpoint - The article discusses the challenges faced by AI painting models in accurately interpreting human instructions and presents Tencent's PromptEnhancer framework as a solution to improve text-image alignment without modifying pre-trained models [2][4][12]. Group 1: Challenges in AI Painting - AI painting models struggle with understanding concise user instructions, leading to inaccuracies in generated images [9][10]. - Common issues include chaotic attribute binding, ineffective negation commands, and failure to comprehend complex spatial relationships [10][11]. Group 2: PromptEnhancer Framework - PromptEnhancer introduces a decoupled prompt optimization framework consisting of two main modules: CoT-based Rewriter and AlignEvaluator [12][14]. - The CoT-based Rewriter mimics human designers by breaking down instructions into core elements, potential ambiguities, and detailed supplements [15][19]. - AlignEvaluator provides a scoring system across 24 key dimensions to accurately identify errors in generated images [20][21]. Group 3: Performance Improvements - Testing on the HunyuanImage 2.1 model shows a 5.1% overall accuracy improvement, with significant gains in complex scene understanding [29]. - Specific dimensions such as "similarity relations" and "counterfactual reasoning" saw accuracy increases of 17.3% and 17.2%, respectively [29]. Group 4: Dataset and Research Support - Tencent's team released a high-quality benchmark dataset containing 6,000 prompts to aid in the training and evaluation of the PromptEnhancer [7][45]. - The dataset covers various complex scenarios, including everyday creative extensions and abstract relationship challenges [46]. Group 5: Future Implications - The advancements brought by PromptEnhancer position it as a critical tool for enhancing AI painting's applicability in professional fields like industrial design and advertising [54][55]. - The framework's ability to optimize instructions without altering model weights allows for broader adaptability across different T2I models [57].