This paper investigates the trade-offs among consistency, diversity, and realism in conditional image generative models, focusing on their potential as world models. The authors analyze state-of-the-art text-to-image (T2I) and image&text-to-image (I-T2I) models, including latent diffusion models (LDM), retrieval-augmented diffusion models (RDM), and neural image compression models like PerCo. They use Pareto fronts to visualize the multi-objective trade-offs between consistency, diversity, and realism. The study finds that while realism and consistency can be improved simultaneously, there is a clear trade-off between realism/consistency and diversity. Earlier models tend to have higher diversity but lower consistency and realism, whereas more recent models excel in consistency and realism but sacrifice diversity. The analysis also reveals geographic disparities in model performance, with some regions showing better consistency and diversity than others. The authors highlight that no single model is optimal for all tasks, and the choice of model should depend on the specific downstream application. They emphasize the importance of Pareto fronts as a tool to evaluate the progress of conditional image generative models toward becoming effective world models. The study also explores the impact of various knobs, such as guidance scale, post-hoc filtering, and retrieval augmentation, on the trade-offs between consistency, diversity, and realism. Overall, the paper provides a comprehensive analysis of the multi-objective trade-offs in conditional image generative models, offering insights into how to select the most appropriate model for different applications. The findings suggest that while recent models have improved in realism and consistency, they may not always be the best choice for tasks requiring high diversity. The study underscores the need for further research to understand and mitigate the trade-offs between these objectives in generative models.This paper investigates the trade-offs among consistency, diversity, and realism in conditional image generative models, focusing on their potential as world models. The authors analyze state-of-the-art text-to-image (T2I) and image&text-to-image (I-T2I) models, including latent diffusion models (LDM), retrieval-augmented diffusion models (RDM), and neural image compression models like PerCo. They use Pareto fronts to visualize the multi-objective trade-offs between consistency, diversity, and realism. The study finds that while realism and consistency can be improved simultaneously, there is a clear trade-off between realism/consistency and diversity. Earlier models tend to have higher diversity but lower consistency and realism, whereas more recent models excel in consistency and realism but sacrifice diversity. The analysis also reveals geographic disparities in model performance, with some regions showing better consistency and diversity than others. The authors highlight that no single model is optimal for all tasks, and the choice of model should depend on the specific downstream application. They emphasize the importance of Pareto fronts as a tool to evaluate the progress of conditional image generative models toward becoming effective world models. The study also explores the impact of various knobs, such as guidance scale, post-hoc filtering, and retrieval augmentation, on the trade-offs between consistency, diversity, and realism. Overall, the paper provides a comprehensive analysis of the multi-objective trade-offs in conditional image generative models, offering insights into how to select the most appropriate model for different applications. The findings suggest that while recent models have improved in realism and consistency, they may not always be the best choice for tasks requiring high diversity. The study underscores the need for further research to understand and mitigate the trade-offs between these objectives in generative models.