Towards Automated Diffusion-Based Synthetic Image Generation for Agricultural News Articles

In modern digital journalism, especially in specific domains such as agriculture, sourcing appropriate and engaging images for daily news articles remains a persistent challenge. This paper explores a novel pipeline that automatically generates high-quality images for news articles using large language models (LLMs) and text-to-image generative models. We curated a dataset of 50 real agricultural news articles and their corresponding editorial images. In the first phase of our study, we compared three frontier LLMs (GPT-4o, Gemini 2.5 Pro, Claude Sonnet 3.7) in generating image prompts based on article title, text, and tags. These prompts were used with Stable Diffusion XL to generate synthetic images, which were then evaluated and ranked by human evaluators. The best-performing LLM was used in the second phase to generate prompts for four different image generation models: Stable Diffusion XL, Stable Diffusion 3.5 Medium, Flux 1 Schnell, and Midjourney 7. Each article thus yielded six images, including one created using a manually crafted Midjourney prompt and one original image selected by the journalist. Human evaluators from the news editorial team ranked the images based on contextual relevance, visual quality, and editorial fit. Our results show that Midjourney-generated images approach the quality of original editorial images, while other open-source models achieved acceptable results in approximately 79% of cases. These findings suggest that AI-generated images can serve as a viable—and in some cases preferable—alternative to manual image selection in news workflows.