Learning Prompt-Level Quality Variance for Cost-Effective Text-to-Image Generation

Text-to-image generation is a multivariable process in which the resulting quality is determined by both the generative model and the input prompt. While previous efforts rely on a single model either by enhancing its capability or by reformulating prompts, we point out that no single model excels at handling all types of tasks, as there exist inter-model and intra-model quality variance induced by the difference in types of prompts. This paper explores the relationship between the generation quality of text-to-image models and the linguistic features of input prompts by measuring the performance of state-of-the-art models using five different prompt datasets each with its distinctive features. Motivated by our empirical observations, we propose a novel approach that assigns each prompt to its best-performing model based on quality prediction. This enables utilizing a diverse set of models each with its expertise and cost, thereby enhancing cost-effectiveness. Evaluation results show that our approach can reduce the total generation cost by 29.25% with comparable or even higher generation quality than using only the single best model.