Eval of 19 ways to crop images using LLMs, Jev and image classification models (workflows-of-ai.com)

A recent exploration into the efficacy of using large language models (LLMs) for image cropping has revealed promising methodologies to automatically identify focal points in older images lacking this metadata. By analyzing 19 selected photos from Pexels, various cropping strategies—including direct prompts for x and y coordinates, rulers, and the Jev decision model—were tested to see how accurately they could locate the subjects in images. Notably, GPT-5.5 demonstrated the best performance, averaging a distance of only 6.2% from the manually marked focal points, while Claude Haiku 5.5 emerged as a cost-effective solution with a median error of 3.8%. This evaluation underscores the significance of improving media library management through AI, addressing a prevalent issue faced by content managers: outdated images that can now be efficiently repurposed with accurate focal points identified by AI. The implications range from enhanced visual storytelling to significant cost savings, especially with models like Jev costing just six cents per 1,000 photos. As the AI/ML community continues to innovate in image processing, such advances could redefine workflows for content creation and optimization.