Imagine asking an AI to sketch a simple circuit diagram or a bar graph showing experimental results. Sounds easy, right? A new research paper titled "ScImage: How Good Are Multimodal Large Language Models at Scientific Text-to-Image Generation?" reveals the surprising truth: while AI has made impressive strides in creating images from text, it still struggles with the precision and unique conventions of scientific visuals. Researchers have developed ScImage, a benchmark designed to test AI's ability to generate scientific images from textual descriptions. This benchmark evaluates three core skills: spatial reasoning (understanding the arrangement of objects), numeric comprehension (accurately representing quantities), and attribute binding (correctly depicting features like color and shape). Five leading AI models, including GPT-4, Llama, and DALL-E, were put to the test using various input languages like English, German, Farsi, and Chinese. The results were intriguing. GPT-4 performed the best, capable of creating decent outputs for simpler diagrams involving single skills like spatial or numeric understanding. However, even GPT-4 faltered when faced with complex prompts that combined multiple skills, often revealing a lack of real-world knowledge or struggling to plan 3D objects on a 2D plane. Surprisingly, models that produced images via code (like Python or TikZ) generally created better "science-styled" visuals than models like DALL-E that directly generate images. This might be because code allows for greater precision and control, something crucial for scientific diagrams. However, these code-based models sometimes produced code with errors that couldn’t be compiled into an image. Another interesting observation was the discrepancy between model types. While code-based models particularly struggled with spatial understanding (like positioning objects correctly), image-based models found numeric comprehension (depicting the correct number of objects) to be the biggest hurdle. The ScImage benchmark highlights the need for further research in this area. While AI can assist scientists with some visualization tasks, it still requires human oversight. Future research should focus on improving AI’s ability to handle complex, multi-dimensional tasks and ensure consistency across diverse scientific fields and languages. As AI models become more integrated into scientific workflows, benchmarks like ScImage are essential for ensuring accuracy, reliability, and ultimately, scientific progress.