Testing whether ChatGPT can tell which way a product photo is rotated (paper) (zenodo.org)

Researchers have introduced BlueVetaUpright, a novel system designed to address the challenge of semantic orientation correction in product photography. While current product-image workflows automate various tasks like cropping and background removal, accurately determining a product's orientation remains largely manual and inconsistent. The study reveals that even advanced vision-language models, including ChatGPT, struggle with this task, often misclassifying 90-degree and 270-degree rotations. In contrast, BlueVetaUpright combines a specialized visual backbone with a trained decision layer to achieve remarkable accuracy—99.02% on standard cardinal rotations and 96.67% across all angles. The significance of this development lies in its potential to enhance automated product photography systems, making them more reliable and efficient. BlueVetaUpright outperformed leading vision-language chat assistants in a comprehensive evaluation, achieving 15.2 percentage points higher accuracy than ChatGPT on a set of 394 photos. This research not only identifies a crucial limitation in existing AI models but also offers a robust solution that could streamline e-commerce and digital marketing by ensuring products are presented in their correct orientation, ultimately improving user experience and decision-making in online shopping.